On Thu, 23 Feb 2023 04:37:49 GMT, Martin Doerr <mdo...@openjdk.org> wrote:
>> Implementation of "Foreign Function & Memory API" for linux on Power (Little >> Endian) according to "Power Architecture 64-Bit ELF V2 ABI Specification". >> >> This PR does not include code for VaList support because it's supposed to >> get removed by [JDK-8299736](https://bugs.openjdk.org/browse/JDK-8299736). >> I've kept the related tests disabled for this platform and throw an >> exception instead. Note that the ABI doesn't precisely specify variable >> argument lists. Instead, it refers to `<stdarg.h>` (2.2.4 Variable Argument >> Lists). >> >> Big Endian support is implemented to some extend, but not complete. E.g. >> structs with size not divisible by 8 are not passed correctly (see >> `useABIv2` in CallArranger.java). Big Endian is excluded by selecting >> `ARCH.equals("ppc64le")` (CABI.java) only. >> >> There's another limitation: This PR only accepts structures with size >> divisible by 4. (An `IllegalArgumentException` gets thrown otherwise.) I >> think arbitrary sizes are not usable on other platforms, either, because >> `SharedUtils.primitiveCarrierForSize` only accepts powers of 2. >> >> The ABI has some tricky corner cases related to HFA (Homogeneous Float >> Aggregate). The same argument may need to get passed in both, a FP reg and a >> GP reg or stack slot (see "no partial DW rule"). This cases are not covered >> by the existing tests. >> >> I had to make changes to shared code and code for other platforms: >> 1. Pass type information when creating `VMStorage` objects from `VMReg`. >> This is needed for the following reasons: >> - PPC64 ABI requires integer types to get extended to 64 bit (also see >> CCallingConventionRequiresIntsAsLongs in existing hotspot code). We need to >> know the type or at least the bit width for that. >> - Floating point load / store instructions need the correct width to select >> between the correct IEEE 754 formats. The register representation in single >> FP registers is always IEEE 754 double precision on PPC64. >> - Big Endian also needs usage of the precise size. Storing 8 Bytes and >> loading 4 Bytes yields different values than on Little Endian! >> 2. It happens that a `NativeMemorySegmentImpl` is used as a raw pointer >> (with byteSize() == 0) while running TestUpcallScope. Hence, existing size >> checks don't work (see MemorySegment.java). As a workaround, I'm just >> skipping the check in this particular case. Please check if this makes sense >> or if there's a better fix (possibly as separate RFE). > > Martin Doerr has updated the pull request incrementally with one additional > commit since the last revision: > > Clean fix for NativeMemorySegmentImpl issue with byteSize 0. > > (I'd be happy to implement the needed changes in shared code if you want, > > since it touches `BindingSpecializer` which is pretty dense) > > FYI: > [master...JornVernee:jdk:I2L](https://github.com/openjdk/jdk/compare/master...JornVernee:jdk:I2L) > (assuming `I2L` has the same semantics as `extsw`). Then just add a > `.cast(int.class, long.class)` wherever currently an `int` is `vmStore`d in > the PPC CallArranger. Correct, `extsw` performs a `I2L` conversion. I had thought about this already, but I think my current implementation is more efficient as it combines register moves with the 64 bit extend. Your proposal would generate separate extend and move instructions, right? ------------- PR: https://git.openjdk.org/jdk/pull/12708
