This is a reland of e588ff10e5
The only change over the original CL is found in JSCreateLowering::AllocateFastLiteral. We now guard against boilerplate values for unboxed double fields that *look* like legitimate initial values, but should really be kHoleNanInt64 instead.
The underlying problem certainly existed before, but an invariant added to LoadElimination in this CL caused a Chromium layout test to fail. The change in this reland is therefore a workaround, the root cause remains to be fixed. Specifically, we find that a pointer to the undefined value oddball is sometimes reinterpreted as a double and assigned as a boilerplate value. @jarin suspects that this stems from in-place map updates.
Original change's description:
> Make LoadElimination aware of const fields (Part 2; stores)
>
> Adds const information to store field accesses and uses it in load elimination
>
> Change-Id: I00765c854c95c955dabd78557463267b95f75eef
> Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1611543
> Reviewed-by: Georg Neis <neis@chromium.org>
> Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
> Commit-Queue: Georg Schmid <gsps@google.com>
> Cr-Commit-Position: refs/heads/master@{#61796}
Change-Id: Ie388754890024a3ca7d10c9d4d7391442655b426
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1630676
Reviewed-by: Georg Neis <neis@chromium.org>
Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Commit-Queue: Georg Schmid <gsps@google.com>
Cr-Commit-Position: refs/heads/master@{#61838}
85f257f4c9