RFR: 8335392: C2 MergeStores: enhanced pointer parsing [v18]

[Shaojin Wen]

On Mon, 4 Nov 2024 11:48:49 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> **Background**
>> I am introducing the `MemPointer`, for enhanced pointer parsing. For now, it replaces the much more limited `ArrayPointer` in `MergeStores` (see https://github.com/openjdk/jdk/pull/16245), but eventually it is supposed to be used widely in optimizations for pointer analysis: adjacency, aliasing, etc. I also plan to refactor the `VPointer` from auto-vectorization with it, and unlock more pointer patterns that way - possibly including scatter/gather.
>> 
>> **Details**
>> 
>> The `MemPointer` decomposes a pointer into the form `pointer = con + sum_i(scale_i * variable_i)` - a linear form with a sum of variables and scale-coefficients, plus some constant offset.
>> 
>> This form allows us to perform aliasing checks - basically we can check if two pointers are always at a constant offset. This allows us to answer many questions, including if two pointers are adjacent. `MergeStores` needs to know if two stores are adjacent, so that we can safely merge them.
>> 
>> More details can be found in the description in `mempointer.hpp`. Please read them when reviewing!
>> 
>> `MemPointer` is more powerful than the previous `ArrayPointer`: the latter only allows arrays, the former also allows native memory accesses, `Unsafe` and `MemorySegement`.
>> 
>> **What this change enables**
>> 
>> Before this change, we only allowed merging stores to arrays, where the store had to have the same type as the array element (`StoreB` on `byte[]`, `StoreI` on `int[]`).
>> 
>> Now we can do:
>> - Merging `Unsafe` stores to array. Including "mismatched size": e.g. `putChar` to `byte[]`.
>> - Merging `Unsafe` stores to native memory.
>> - Merging `MemorySegment`: with array, native, ByteBuffer backing types.
>>   - However: there is still some problem with RangeCheck smearing (a type of RC elimination) for the examples I have tried. Without RC's smeared, we can only ever merge 2 neighbouring stores. I hope we can improve this with better RangeCheck smearing. `MemorySegment` introduce `checkIndexL`, the long-variant of the RangeCheck. Normal array accesses only use the equivalent of `checkIndex`, the int-variant that we already optimize away much better.
>> 
>> **Dealing with Overflows**
>> 
>> We have to be very careful with overflows when dealing with pointers. For this, I introduced a `NoOverflowInt`. It allows us to do "normal" int operations on it, and tracks if there was ever an overflow. This way, we can do all overflow checks implicitly, and do not clutter the code with overflow-check...
>
> Emanuel Peter has updated the pull request incrementally with one additional commit since the last revision:
> 
>   more changes for Christian

If it is not provided in the release image, users need to find the source code of the current version of JDK to build the fastdebug image to analyze whether the MergeStore optimization of a certain code works.

I can understand that MergeStore may still need to be improved, so it cannot be used as a product feature, but this is a useful optimization and I hope it can be provided in the product eventually.

I hope that TraceMergeStore can eventually be used in the release image like `PrintInlining` and become a tool for performance optimizers.

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

PR Comment: https://git.openjdk.org/jdk/pull/19970#issuecomment-2461092539