On Fri, 22 Nov 2024 14:31:23 GMT, Jorn Vernee <jver...@openjdk.org> wrote:
> Allow `captureCallState` and `critical(true)` linker options to be combined. > This allows passing a Java array to capture call state. > > One caveat is that the linker expects the memory to be aligned, which means > that at least an `int[]` has to be used (i.e. `byte[]` will no work). > > This patch contains two implementations: one for the linkers that use > `CallingSequenceBuilder`. That one is quite straight-forward, as we can just > mimic what we already do for other memory segment arguments, but also for the > capture state segment. i.e. split it into base and offset, and pass that down > to our downcall stub. The stub will then add the offset and oop together, and > pass use the resulting address to write to. > > The other implementation is for the fallback linker. This handles the capture > state a little differently, but essentially currently just passes the native > address to the back end for the native code to write the captured state into. > I've just added another heap base parameter for that capture state segment to > the back end, which is then turned into a native address using JNI's > `GetPrimitiveArrayCritical`, similarly to what we do for other heap segments. > > Testing: `jdk_foreign` test suite. src/java.base/share/classes/jdk/internal/foreign/abi/NativeEntryPoint.java line 91: > 89: int checkIdx = 1; > 90: if ((needsReturnBuffer && methodType.parameterType(checkIdx++) != > long.class) > 91: || (savedValueMask != 0 && Maybe this routine is getting too complex and it would be better to split the checks (and add some comments) ? E.g. we need to check that if there's a return buffer, a certain low-level argument is `long` and, if there's need to capture state, we either have `long`, or an `Object,long` pair. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/22327#discussion_r1861153265
