> Flat accesses to a stable value can be expanded in a non-atomic way if the
> stable field is already initialized since they are read-only at this point.
> That allows to make more optimizations, and in particular to replace the
> access by a constant if it's known at compilation time.
>
> There are 2 cases. First, flat stable non-array field. In this case, the
> value is known to be stable if the value is non-null, that is if the
> null-marker of the said field is 1. If we can find that the null marker has a
> constant value that is non zero, we expand non-atomically. That is done by
> finding the field we are trying to get from the offset. From the field, we
> can find the offset of the null marker, and then the null marker `ciField`,
> allowing to fetch its value if the holder is known to be a constant oop.
>
> The other case is stable flat array. In this case, we need to find index of
> the containing element of the array, then with the offset, we can find the
> field we are trying to get. Finding the null marker here is a bit more
> tricky. Let's say we have
>
> value class MyValue {
> int x;
> }
> class C {
> MyValue v; // assumed flat.
> }
>
> the null marker for `v` is somewhat a field of `C`, as well as `v.x`. So I
> can just use `field_value` to get the null marker. But in `MyValue[]`, there
> isn't a single field for the null marker, but one "field" for each cell of
> the array, and there isn't a nice containing type in which it lives. The only
> way to get each piece of the array is by index (or offset). So, I needed
> specialized functions to access the null marker of a cell given the
> index/offset.
>
> I also had to implement of bit of accessors for `ciFlatArray`. First,
> implement `element_value_by_offset` in `ciFlatArray` since the implementation
> of `ciArray` (super-class) was used, which computes the index from the
> provided offset, assuming a size of elements that doesn't take flattening
> into account as it used only the basic type, and not the layout helper. But
> also helpers to get a field of the flattened value class in a cell, to allow
> constant folding to get the constant value in an array.
>
> The last part of the puzzle, necessary to make constant folding happen (see
> `Type::make_constant_from_field`), is to say that a field of a flattened
> inline type field is constant if the containing field if constant. In the
> words of the previous example, that means `x` is constant in `C` if `v` is
> strict and final (already there), or if `v` is constant itself. That matches
> what we do in `void ciField::i...
Marc Chevalier has updated the pull request with a new target base due to a
merge or a rebase. The pull request now contains 16 commits:
- Merge remote-tracking branch 'origin/lworld' into JDK-8372700
- review
- Solving conflict with myself: Merge remote-tracking branch 'lworld' into
JDK-8372700
- Un-problemlist
- Merge remote-tracking branch 'origin/lworld' into JDK-8372700
- Cleanup
- Cleanup
- Fix header size
- damn is_null_or_zero assumes is_valid
- Merge remote-tracking branch 'origin/lworld' into JDK-8372700
- ... and 6 more: https://git.openjdk.org/valhalla/compare/69399ced...b64ff8ce
-------------
Changes: https://git.openjdk.org/valhalla/pull/1826/files
Webrev: https://webrevs.openjdk.org/?repo=valhalla&pr=1826&range=02
Stats: 301 lines in 20 files changed: 280 ins; 7 del; 14 mod
Patch: https://git.openjdk.org/valhalla/pull/1826.diff
Fetch: git fetch https://git.openjdk.org/valhalla.git pull/1826/head:pull/1826
PR: https://git.openjdk.org/valhalla/pull/1826
