On Tue, 11 Jun 2024 11:35:28 GMT, Shaojin Wen <d...@openjdk.org> wrote:
>> After PR https://github.com/openjdk/jdk/pull/16245, C2 optimizes stores into >> primitive arrays by combining values into larger stores. >> >> This PR rewrites the code of appendNull and append(boolean) methods so that >> these two methods can be optimized by C2. > > Shaojin Wen has updated the pull request incrementally with one additional > commit since the last revision: > > revert Nice, thanks for the extra logs. Ok, the optimization is at least working. Maybe not always fully. In 1.2 we see that only 2 stores were merged, those from lines 71 and 72. So that would be a potential loss of performance. That could be interesting to debug. Can you create a stand-alone reproducer, so just a single java file, where the code looks as similar as here, an where it produces the same kind of `TraceMergeStores` output? I could look into that. Also: do you expect only the "not" `isLatin1()` case to merge the stores, or also in the `isLatin1()` case? It would be interesting here to extract the relevant (and only the relevant) assembly code. There we would be looking for the stores. These would be byte-stores if not merged, and fewer but larger stores if merged. Since the assembly is annotated with the methods, you can find it relatively easily. You will be looking for `mov` kinds of instructions that store register values into addresses. You can easily learn more about assembly and `mov` instructions online. General explanations of `mov` instructions: https://c9x.me/x86/html/file_module_x86_id_176.html https://www.felixcloutier.com/x86/mov About the suffixes, i.e. `movq` vs `movl` etc. https://stackoverflow.com/questions/67690999/assembly-do-we-need-the-endings I hope this helps ;) ------------- PR Comment: https://git.openjdk.org/jdk/pull/19626#issuecomment-2162114560
