RFR: 8267532: Try/catch block not optimized as expected

[Jorn Vernee]

On Mon, 30 Oct 2023 14:10:33 GMT, Jorn Vernee <jvernee at openjdk.org> wrote:

> The issue is essentially that for the Java try-with-resource construct, javac generates multiple calls to `close(`) the resource. One of those calls is inside the hidden exception handler of the try block. The issue for us is that typically the exception handler is never entered (since no exception is thrown), however we don't profile exception handlers at the moment, so the block is not pruned. C2 doesn't inline the `close()` call in the handler due to low call site frequency. As a result, the receiver of that call escapes and can not be scalar replaced, which then leads to a loss in performance.
> 
> There has been some discussion on the JBS issue that this could be fixed by profiling catch blocks. And another suggestion that partial escape analysis could help here to prevent the object from escaping. But, I think there are other benefits to being able to prune dead catch blocks, such as general reduction in code size, and other optimizations being possible by dead code being eliminated. So, I've implemented catch block profiling + pruning in this patch.
> 
> The implementation is essentially very straightforward: we allocate an extra bit of profiling data for each
> exception handler of a method in the `MethodData` for that method (which holds all the profiling
> data). Then when looking up the exception handler after an exception is thrown, we mark the
> exception handler as entered. When C2 parses the exception handler block, and it sees that it has
> never been entered, we emit an uncommon trap instead.
> 
> I've also cleaned up the handling of profiling data sections a bit. After adding the extra section of data to MethodData, I was seeing several crashes when ciMethodData was used. The underlying issue seemed to be that the offset of the parameter data was computed based on the total data size - parameter data size (which doesn't work if we add an additional section for exception handler data). I've re-written the code around this a bit to try and prevent issues in the future. Both MethodData and ciMethodData now track offsets of parameter data and exception handler data, and the size of the each data section is derived from the offsets.
> 
> Finally, there was an assert firing in `freeze_internal` in `continuationFreezeThaw.cpp`:
> 
>     assert(monitors_on_stack(current) == ((current->held_monitor_count() - current->jni_monitor_count()) > 0),
>          "Held monitor count and locks on stack invariant: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count...

src/hotspot/share/runtime/sharedRuntime.cpp line 681:

> 679: // for given exception
> 680: // Note that the implementation of this method assumes it's only called when an exception has actually occured
> 681: address SharedRuntime::compute_compiled_exc_handler(CompiledMethod* cm, address ret_pc, Handle& exception,

One thing of note for this function: We don't look up the exception handler bci for JVMCI compiled methods, so I've not added any profiling in the case of JVMCI (see the `#if INCLUDE_JVMCI` block at the start of the function).

This means that when using a JVMCI compiler, exception handlers might appear as untaken, when they are actually taken. This should be fine since the profiling information is currently only used by C2. But, if a JVMCI compiler wants to start using the profiling information e.g. to prune dead exception handlers as well, then profiling needs to be implement first.

Please let me know if this is okay.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/16416#discussion_r1378839166