RFR: 8325467: Support methods with many arguments in C2 [v4]

[Daniel Lundén]

On Tue, 20 Aug 2024 14:32:25 GMT, Daniel Lundén <dlunden at openjdk.org> wrote:

>> If a method has a large number of parameters, we currently bail out from C2 compilation.
>> 
>> ### Changeset
>> 
>> Allowing C2 compilation of methods with a large number of parameters requires fundamental changes to the register mask data structure, used in many places in C2. In particular, register masks currently have a statically determined size and cannot represent arbitrary numbers of stack slots. This is needed if we want to compile methods with arbitrary numbers of parameters. Register mask operations are present in performance-sensitive parts of C2, which further complicates changes.
>> 
>> Changes:
>> - Add functionality to dynamically grow/extend register masks. I experimented with a number of design choices to achieve this. To keep the common case (normal number of method parameters) quick and also to avoid more intrusive changes to the current `RegMask` interface, I decided to leave the "base" statically allocated memory for masks unchanged and only use dynamically allocated memory in the rare cases where it is needed.
>> - Generalize the "chunk"-logic from `PhaseChaitin::Select()` to allow arbitrary-sized chunks, and also move most of the logic into register mask methods to separate concerns and to make the `PhaseChaitin::Select()` code more readable.
>> - Remove all `can_represent` checks and bailouts.
>> - Performance tuning. A particularly important change is the early-exit optimization in `RegMask::overlap`, used in the performance-sensitive method `PhaseChaitin::interfere_with_live`.
>> - Add a new test case `TestManyMethodArguments.java` and extend an old test `TestNestedSynchronize.java`.
>> 
>> ### Testing
>> 
>> - [GitHub Actions](https://github.com/dlunde/jdk/actions/runs/10178060450)
>> - `tier1` to `tier4` (and additional Oracle-internal testing) on Windows x64, Linux x64, Linux aarch64, macOS x64, and macOS aarch64.
>> - Standard performance benchmarking. No observed conclusive overall performance degradation/improvement.
>> - Specific benchmarking of C2 compilation time. The changes increase C2 compilation time by, approximately and on average, 1% for methods that could also be compiled before this changeset (see the figure below). The reason for the degradation is further checks required in performance-sensitive code (in particular `PhaseChaitin::remove_bound_register_from_interfering_live_ranges`). I have tried optimizing in various ways, but changes I found that lead to improvement also lead to less readable code (and are, in my opinion, no...
>
> Daniel Lundén has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Update after Roberto's comments and suggestions

I explored the possibility of calculating a stacic upper bound for register mask size given that parameters must fit in 255 32-bit words (from the JVM spec). My conclusion is that the required increase in static register mask size is too expensive, and that we should proceed with my current solution.

**Details**
In C2, register masks must be capable of representing incoming arguments (up to 255) as well as the maximum number of outgoing arguments among all function calls in the compiled method (also up to 255). Additionally, arguments are 64-bit aligned in the frame, which doubles the number of required register mask bits.

Example on my x64 machine (1 word = 32 bits here):
- Registers require at minimum 18 words in register masks.
- We currently add 4 words for representing arguments, locks, and some other stack locations, so in total 22 words.
- To ensure we can fit 255 incoming arguments in the mask, we need 255 * 2 bits = 16 words.
- To ensure we can fit 255 outgoing arguments in the mask, we need 255 * 2 bits = 16 words.

That is, we go from 18 + 4 = 22 words to at least 18 + 16 + 16 = 50 words, only taking incoming and outgoing arguments into account. Performance experiments indicate a 3.3% C2 compilation speed degradation (compared to 1% for the solution in this PR).

-------------

PR Comment: https://git.openjdk.org/jdk/pull/20404#issuecomment-2312905150