RFR: 8352075: Perf regression accessing fields [v21]

[Ioi Lam]

On Thu, 5 Jun 2025 05:17:48 GMT, Ioi Lam <iklam at openjdk.org> wrote:

>> Radim Vansa has updated the pull request incrementally with three additional commits since the last revision:
>> 
>>  - Moved jtreg test
>>  - Improved documentation
>>  - Fix coding style (asterisk placement)
>
> I have written a POC that shows that the table must be sorted again when dumping a dynamic CDS archive. See https://github.com/iklam/jdk/commit/dcd53ebaeab7b38be02aa5b896ce9e449a45418f
> 
> Explanations are in [here](https://github.com/iklam/jdk/commit/dcd53ebaeab7b38be02aa5b896ce9e449a45418f#diff-fd7608607ecf305bb3535b500bff5d53ec216d2da25e3bad1a1d699f56b09283R199)
> 
> I will create an RFE for the JDK mainline that adds built-in debugging support for the `(oldSym > newSym_orig)` condition as describe in the POC. Please wait for that before integrating this PR. I can help you write the code for re-sorting the tables.

> As @iklam was saying in above, the table won't work for dynamic dumping for CDS and fails all these tests.
> 
> ```
>  Array<u1>* FieldInfoStream::create_search_table(ConstantPool* cp, const Array<u1>* fis, ClassLoaderData* loader_data, TRAPS) {
> +  if (CDSConfig::is_dumping_dynamic_archive()) {
> +    // We cannot call validate_search_table. The _fieldinfo_search_table should be sorted by "requested" addresses,
> +    // but validate_search_table will be getting Symbol* addresses from _constants, which has "buffered" addresses.
> +    //
> +    // For background, see new comments inside allocate_node_impl in symbolTable.cpp
> +    return nullptr;
> +  }
> +
> ```
> 
> This fixes these tests.

This fix looks good to me. It also obviates the need to re-sort the table for dynamic CDS dump. I am OK with this. We can implement re-sorting for the dynamic CDS archives in a separate RFE if desired.

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

PR Comment: https://git.openjdk.org/jdk/pull/24847#issuecomment-2946819592