RFR: 8331341: secondary_super_cache does not scale well: C1 and interpreter [v9]

[Vladimir Ivanov]

On Fri, 26 Jul 2024 15:13:06 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> This patch expands the use of a hash table for secondary superclasses
>> to the interpreter, C1, and runtime. It also adds a C2 implementation
>> of hashed lookup in cases where the superclass isn't known at compile
>> time.
>> 
>> HotSpot shared runtime
>> ----------------------
>> 
>> Building hashed secondary tables is now unconditional. It takes very
>> little time, and now that the shared runtime always has the tables, it
>> might as well take advantage of them. The shared code is easier to
>> follow now, I think.
>> 
>> There might be a performance issue with x86-64 in that we build
>> HotSpot for a default x86-64 target that does not support popcount.
>> This means that HotSpot C++ runtime on x86 always uses a software
>> emulation for popcount, even though the vast majority of machines made
>> for the past 20 years can do popcount in a single instruction. It
>> wouldn't be terribly hard to do something about that.
>> 
>> Having said that, the software popcount is really not bad.
>> 
>> x86
>> ---
>> 
>> x86 is rather tricky, because we still support
>> `-XX:-UseSecondarySupersTable` and `-XX:+UseSecondarySupersCache`, as
>> well as 32- and 64-bit ports. There's some further complication in
>> that only `RCX` can be used as a shift count, so there's some register
>> shuffling to do. All of this makes the logic in macroAssembler_x86.cpp
>> rather gnarly, with multiple levels of conditionals at compile time
>> and runtime.
>> 
>> AArch64
>> -------
>> 
>> AArch64 is considerably more straightforward. We always have a
>> popcount instruction and (thankfully) no 32-bit code to worry about.
>> 
>> Generally
>> ---------
>> 
>> I would dearly love simply to rip out the "old" secondary supers cache
>> support, but I've left it in just in case someone has a performance
>> regression.
>> 
>> The versions of `MacroAssembler::lookup_secondary_supers_table` that
>> work with variable superclasses don't take a fixed set of temp
>> registers, and neither do they call out to to a slow path subroutine.
>> Instead, the slow patch is expanded inline.
>> 
>> I don't think this is necessarily bad. Apart from the very rare cases
>> where C2 can't determine the superclass to search for at compile time,
>> this code is only used for generating stubs, and it seemed to me
>> ridiculous to have stubs calling other stubs.
>> 
>> I've followed the guidance from @iwanowww not to obsess too much about
>> the performance of C1-compiled secondary supers lookups, and to prefer
>> simplicity over absolute performance. Nonetheless, this i...
>
> Andrew Haley has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Fix test failure

Oh, it comes as a surprise to me... I was under impression that the first thing hand-coded assembly variants do is check  for `bitmap != SECONDARY_SUPERS_BITMAP_FULL`. At least, it was my recollection from working on [JDK-8180450](https://bugs.openjdk.org/browse/JDK-8180450). (And the initial version of the patch     with the check in `Klass::lookup_secondary_supers_table()` and  `_bitmap(SECONDARY_SUPERS_BITMAP_FULL)` only reassured me that's still the case.)

> The root cause of this bug is that we're not maintaining the invariant popcount(bitmap) == secondary_supers()->length().

The invariant holds only when `bitmap != SECONDARY_SUPERS_BITMAP_FULL`. It does help that even in case of non-hashed `secondary_supers` list `secondary_supers()->length() >= popcount(bitmap)`, but initial probing becomes much less efficient (a random probe followed by a full linear pass over secondary supers list).

Alternatively, all table lookups can be adjusted to start with `bitmap != SECONDARY_SUPERS_BITMAP_FULL` checks before probing the table. It does add a branch on the fast path (and slightly increases inlined snippet code size), but the branch is highly predictable and works on a value we need anyway. So, I would be surprised to see any performance effects from it. IMO it's easier to reason and more flexible: `SECONDARY_SUPERS_BITMAP_FULL == bitmap` simply disables all table lookups and unconditionally falls back to linear search.

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

PR Comment: https://git.openjdk.org/jdk/pull/19989#issuecomment-2253281836