AArch64: Fix getArgPointer() for FFI upcall #16332
Conversation
knn-k
added
comp:vm
project:panama
|
Jenkins test sanity alinux64,amac jdk19 |
knn-k
force-pushed
the
aarch64upcallFix2
branch
from
bffc313
to
52c5c03
Compare
|
Jenkins test sanity alinux64,amac jdk19 |
|
This PR fixes the failures with TestUpcallHighArity.java on AArch64 in #16237. |
|
FYI. @ChengJin01 |
This commit fixes getArgPointer() for AArch64 for FFI upcall. The function needs to maintain two offsets in the stack: one for the arguments stored by the upcall thunk, another one for the arguments stored in the stack by the caller. Signed-off-by: KONNO Kazuhiro <konno@jp.ibm.com>
knn-k
force-pushed
the
aarch64upcallFix2
branch
from
52c5c03
to
ad8d0e5
Compare
|
Jenkins test sanity alinux64,amac jdk19 |
|
I ran the following tests with this fix on Linux and macOS manually. TestUpcallAsync.java TestUpcallStack.java failed only on macOS, and #16336 tracks it. All other tests were successful. |
|
maintaining two stack-pointer(s) sounds more complicated than necessary. it can be done relatively simply: saving-frame is of fixed size because you can know the max stack saving area needed for all possibly in-register args. at the same time, original passed-in-stack arg pointer can be calculated according to that fixed size and arg's offset together. z (and to a lesser degree, p where the ABI dictates the stack saving area either none or all) implemented this way. |
|
your fix did implement that way. implementation details are slightly different. LGTM ... approved! |
This commit fixes getArgPointer() for AArch64 for FFI upcall.
The function needs to maintain two offsets in the stack: one for the arguments stored by the upcall thunk, another one for the arguments stored in the stack by the caller.
Fixes: #16237
Signed-off-by: KONNO Kazuhiro konno@jp.ibm.com