This adds a new flag --modify-field-representation-inplace (enabled by
default), which lets the runtime perform field representation changes
for Smi to Tagged or for HeapObject to Tagged in-place instead of
creating new maps and marking the previous map tree as deprecated.
That means we create (a lot) fewer Maps and DescriptorArrays in the
beginning and also need to self-heal fewer objects later (migrating
off the deprecated maps). In TurboFan we just take the "field owner
dependency" whenever we use the field representation, which is very
similar to what we already do for the field types. That means if we
change the representation of a field that we used in optimized code,
we will simply deoptimize that code and have TurboFan potentially
later optimize it again with the new field representation.
On the Speedometer2/ElmJS-TodoMVC test, this reduces the total execution
time from around 415ms to around 352ms, which corresponds to a **15%**
improvement. The overall Speedometer2 score improves from around 74.1
to around 78.3 (on local runs with content_shell), corresponding to a
**5.6%** improvement here. 🎉
On the CNN desktop browsing story, it seems that we reduce map space
utilization/fragmentation by about 4-5%. But since we allocate a lot
less (fewer Maps and DescriptorArrays) we also significantly change
the GC timing, which heavily influences the results here. So take this
with a grain of salt. 🤷
Note: For Double fields, this doesn't change anything, meaning they
still create new maps and deprecate the previous map trees.
Bug: v8:8749, v8:8865, v8:9114
Change-Id: Ibd70efcb59be982863905663dbfaa89aa5b31e14
Cq-Include-Trybots: luci.chromium.try:linux-rel,win7-rel
Doc: http://bit.ly/v8-in-place-field-representation-changes
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1565891
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Commit-Queue: Igor Sheludko <ishell@chromium.org>
Auto-Submit: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Igor Sheludko <ishell@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60822}
This reverts commit 1416d5a565.
Reason for revert: blocks roll https://chromium-review.googlesource.com/c/chromium/src/+/1564550
Original change's description:
> [map] Support in-place field representation changes.
>
> This adds a new flag --modify-field-representation-inplace (enabled by
> default), which lets the runtime perform field representation changes
> for Smi to Tagged or for HeapObject to Tagged in-place instead of
> creating new maps and marking the previous map tree as deprecated.
>
> That means we create (a lot) fewer Maps and DescriptorArrays in the
> beginning and also need to self-heal fewer objects later (migrating
> off the deprecated maps). In TurboFan we just take the "field owner
> dependency" whenever we use the field representation, which is very
> similar to what we already do for the field types. That means if we
> change the representation of a field that we used in optimized code,
> we will simply deoptimize that code and have TurboFan potentially
> later optimize it again with the new field representation.
>
> On the Speedometer2/ElmJS-TodoMVC test, this reduces the total execution
> time from around 415ms to around 352ms, which corresponds to a **15%**
> improvement. The overall Speedometer2 score improves from around 74.1
> to around 78.3 (on local runs with content_shell), corresponding to a
> **5.6%** improvement here. 🎉
>
> On the CNN desktop browsing story, it seems that we reduce map space
> utilization/fragmentation by about 4-5%. But since we allocate a lot
> less (fewer Maps and DescriptorArrays) we also significantly change
> the GC timing, which heavily influences the results here. So take this
> with a grain of salt. 🤷♂️
>
> Note: For Double fields, this doesn't change anything, meaning they
> still create new maps and deprecate the previous map trees.
>
> Bug: v8:8749, v8:8865, v8:9114
> Change-Id: I694a53f87ae5caeb868fd98a21809b66d4297d35
> Cq-Include-Trybots: luci.chromium.try:linux-blink-rel
> Doc: http://bit.ly/v8-in-place-field-representation-changes
> Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1561132
> Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
> Reviewed-by: Igor Sheludko <ishell@chromium.org>
> Reviewed-by: Jaroslav Sevcik <jarin@chromium.org>
> Reviewed-by: Georg Neis <neis@chromium.org>
> Cr-Commit-Position: refs/heads/master@{#60764}
TBR=jarin@chromium.org,neis@chromium.org,ishell@chromium.org,bmeurer@chromium.org,verwaest@chromium.org
# Not skipping CQ checks because original CL landed > 1 day ago.
Bug: v8:8749, v8:8865, v8:9114
Change-Id: I666975d08d51bbe7ab4faec9428b9a1f88e9b322
Cq-Include-Trybots: luci.chromium.try:linux-blink-rel
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1564208
Reviewed-by: Michael Hablich <hablich@chromium.org>
Commit-Queue: Michael Hablich <hablich@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60807}
This adds a new flag --modify-field-representation-inplace (enabled by
default), which lets the runtime perform field representation changes
for Smi to Tagged or for HeapObject to Tagged in-place instead of
creating new maps and marking the previous map tree as deprecated.
That means we create (a lot) fewer Maps and DescriptorArrays in the
beginning and also need to self-heal fewer objects later (migrating
off the deprecated maps). In TurboFan we just take the "field owner
dependency" whenever we use the field representation, which is very
similar to what we already do for the field types. That means if we
change the representation of a field that we used in optimized code,
we will simply deoptimize that code and have TurboFan potentially
later optimize it again with the new field representation.
On the Speedometer2/ElmJS-TodoMVC test, this reduces the total execution
time from around 415ms to around 352ms, which corresponds to a **15%**
improvement. The overall Speedometer2 score improves from around 74.1
to around 78.3 (on local runs with content_shell), corresponding to a
**5.6%** improvement here. 🎉
On the CNN desktop browsing story, it seems that we reduce map space
utilization/fragmentation by about 4-5%. But since we allocate a lot
less (fewer Maps and DescriptorArrays) we also significantly change
the GC timing, which heavily influences the results here. So take this
with a grain of salt. 🤷♂️
Note: For Double fields, this doesn't change anything, meaning they
still create new maps and deprecate the previous map trees.
Bug: v8:8749, v8:8865, v8:9114
Change-Id: I694a53f87ae5caeb868fd98a21809b66d4297d35
Cq-Include-Trybots: luci.chromium.try:linux-blink-rel
Doc: http://bit.ly/v8-in-place-field-representation-changes
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1561132
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Igor Sheludko <ishell@chromium.org>
Reviewed-by: Jaroslav Sevcik <jarin@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60764}
This CL includes runtime and IC parts of the tracking. It is controlled by
compile-time flag FLAG_constant_field_tracking and currently disabled.
Transition from kConst to kMutable still involves map deprecation.
BUG=v8:5495
Review-Url: https://codereview.chromium.org/2598543003
Cr-Commit-Position: refs/heads/master@{#43081}
