Implement boxing/unboxing during VectorAPIExpansion under an option #20685
Conversation
gita-omr
commented
Nov 27, 2024
gita-omr
commented
- introduce new category of temps: vec_len_boxed_unknown
- build temp and whole webs as before
- wherever a temp was invalidated before assign vec_len_boxed_unknown instead
- store vectorization info for calls per node vs. per symbol reference for temps
- always vectorize nodes that are calls to vector API methods, even if input vectors need to be unboxed or result needs to be boxed
- since boxing/unboxing is not supported for all types, have an extra pass just before transformation and invalidate whole web if it contains unsupported boxing/unboxing
- transformation happens as before, except that we first check if children of non-vectorized node need to be boxed or children of vectorized node need to be unboxed, insert boxing/unboxing correspondingly
- this prototype is only enabled under -Xjit:enableVectorAPIBoxing option and the default behaviour shoulld be preserved
gita-omr
added this to the Release 0.49 (Java 8, 11, 17, 21, 23) January refresh milestone
gita-omr
changed the title
|
Noticed a bug.. |
|
Fixed some bugs and addressed second round of comments. openjdk sanity passes (with and without the option). However, I had to disable boxing of Mask objects due to a failure in that bucket. Debugging now. |
gita-omr
force-pushed
the
vapi_boxing_pr_2
branch
from
1e69da9 to
bce9f68
Compare
|
Added improvements and addressed latest comments except for the doxygen comments. |
gita-omr
force-pushed
the
vapi_boxing_pr_2
branch
from
bce9f68 to
7f5af92
Compare
gita-omr
changed the title
|
Squashed and removed WIP. |
- introduce new category of temps: vec_len_boxed_unknown - build temp and whole webs as before - wherever a temp was invalidated before assign vec_len_boxed_unknown instead - store vectorization info for calls per node vs. per symbol reference for temps - always vectorize nodes that are calls to vector API methods, even if input vectors need to be unboxed or result needs to be boxed - since boxing/unboxing is not supported for all types, have an extra pass just before transformation and invalidate whole web if it contains unsupported boxing/unboxing - transformation happens as before, except that we first check if children of non-vectorized node need to be boxed or children of vectorized node need to be unboxed, insert boxing/unboxing correspondingly - this prototype is only enabled under -Xjit:enableVectorAPIBoxing option and the default behaviour shoulld be preserved
gita-omr
force-pushed
the
vapi_boxing_pr_2
branch
from
7f5af92 to
9db7f57
Compare
|
Addressed latest comment. |
|
Jenkins test sanity alinux64,plinux,xlinux,zlinux jdk21 |
|
s390 failures seem to be unrelated. |
|
Jenkins test sanity alinux64,plinux,xlinux,zlinux jdk21 |
|
A test job on xlinux failed with the message below: |
|
Jenkins test sanity.functional xlinux jdk21 |
|
Test failures on plinux and zlinux are related to CRIU. They don't seem to be caused by this PR. Waiting for xlinux results. plinux: zlinux: |
| switch (type) | ||
| { | ||
| case TR::Float: | ||
| j9class = vm->floatArrayClass; |
There was a problem hiding this comment.
FYI: Accessing J9JavaVM directly in the optimizer will break JITServer functionality. Typically a frontend function needs to be created to fetch the appropriate field and that function needs to be overridden at the server frontend. PR #20592 shows an example
| newObject->setSymbolReference(comp()->getSymRefTab()->findOrCreateNewObjectSymbolRef(comp()->getMethodSymbol())); | ||
|
|
||
| TR_J9VMBase *fej9 = comp()->fej9(); | ||
| TR::VMAccessCriticalSection getClassFromSignature(fej9); |
There was a problem hiding this comment.
Why is this VMaccessCriticalSection needed here? Typically such sections access objects and we want to prevent GC moving those objects around while we inspect them. Such pieces of code cannot be executed at the server; the server will send a message to the client and ask it to do the work on its behalf.
| if (_trace) | ||
| traceMsg(comp(), "Boxed: %dth child of node %p into %p\n", i, node, newObject); | ||
|
|
||
| if (TR::Options::getVerboseOption(TR_VerboseVectorAPI)) |
There was a problem hiding this comment.
I would prefer the optimizer to use compilation logs instead of vlog. Vlog is typically used for runtime. Verbose options are global and they are not passed to the JITServer If the method is compiled remotely, these messages will not appear.
