Add profiler ELT test #39550
Add profiler ELT test #39550
Conversation
|
Tagging subscribers to this area: @tommcdon |
|
Thanks! This will be interesting when run conjunction with gc stress. Should happen in our overnight testing, once this is merged. |
davmason
force-pushed
the
elt_tests
branch
2 times, most recently
from
2c192c2
to
d005b1c
Compare
|
There are some bugs in the runtime's slow path ELT assembly stubs on linux. Not surprising given that we haven't ever been testing them. I just pushed some changes to add more diagnostics, I am going to go through and root cause them. |
davmason
force-pushed
the
elt_tests
branch
2 times, most recently
from
d874005
to
bdd8471
Compare
davmason
force-pushed
the
elt_tests
branch
2 times, most recently
from
a83e7b1
to
4714690
Compare
|
@noahfalk the remaining failures are an infrastructure issue, this is ready to review now @AndyAyersMS do you or anyone else from the jit team want to take a look at the assembly stub changes I made? I found the following issues:
If there are other interesting corner cases to test, please let me know. |
|
cc @dotnet/jit-contrib |
|
@jashook this is the PR we talked about in standup, didn't realize you weren't part of the dotnet-diag group |
That is fixed now right or is there still an issue? |
There was a problem hiding this comment.
LGTM, though mostly I am relying on your test cases rather than my ability spot suspicious assembly code : )
As a broader issue if we've had these long standing bugs and nobody is complaining we should keep an eye out for feedback whether these APIs matter to the profiler authors. Its not clear to me that they do.
| @@ -144,7 +144,7 @@ Stack for the above call will look as follows (stack growing downwards): | |||
| Thread::VirtualUnwindCallFrame(&ctx); | |||
|
|
|||
| // add the prespill register(r0-r3) size to get the stack pointer of previous function | |||
| _ASSERTE(pData->profiledSp == (void*)(ctx.Sp - 4*4)); | |||
| _ASSERTE(pData->profiledSp == (void*)(ctx.Sp - 4*4) || pData->profiledSp == (void*)(ctx.Sp - 6*4)); | |||
There was a problem hiding this comment.
The comment above the assert doesn't lead me to expect the 6 dword case, update the comment?
|
The CI legs marked as in progress are actually complete, going to merge. |
Fix the following issues:
On amd64 linux we didn't save and restore the xmm registers, and didn't handle enregistered 16 bytes structs as return values
On arm we didn't save and restore the floating point registers (I made the linux arm helpers match the windows arm helpers)
On arm64 we didn't handle 16 byte enregistered structs as return values
And add tests
ghost
locked as resolved and limited conversation to collaborators
As mentioned in #39335 we currently have no ELT hook coverage in the runtime repo.
cc @AndyAyersMS