Replace CodeManState with cache in EECodeInfo for x86. #114170
Conversation
ghost
added
the
dotnet-policy-service
bot
added
the
community-contribution
|
Tagging subscribers to this area: @mangod9 |
|
This is not an universal win. It makes On other platforms it saves an unused space. There seem to be no downsides. It gets rid of one x86 quirk and makes it easier to reuse the cache in Thoughts? (I am aware of the build issues due to mismatching offsets; not worth resolving if this whole direction gets abandoned) |
|
LGTM! I like that it gets rid of the unused cache on !x86. |
Thanks for having a look. I made some progress on this locally and it helped me improve the performance on x86+funclet platforms. I'll clean up the PR soon. |
filipnavara
force-pushed
the
no-codemanstate
branch
from
fbe883d to
16b67b1
Compare
filipnavara
changed the title
On all other platforms CodeManState was unused and taking space in StackFrameIterator. Optimize some code paths to cache GC info early or decode the method size without decoding the whole table. Move PCTAddr (Eip pointer) into KNONVOLATILE_CONTEXT_POINTERS_EX to avoid recomputing it when switching between callee and caller contexts.
filipnavara
force-pushed
the
no-codemanstate
branch
from
ca82fc0 to
363f5f7
Compare
| inline TADDR GetRegdisplayPCTAddr(REGDISPLAY *display) | ||
| { | ||
| return display->PCTAddr; | ||
| } | ||
|
|
||
| inline void SetRegdisplayPCTAddr(REGDISPLAY *display, TADDR addr) |
There was a problem hiding this comment.
I initially tried to optimize how the PCTAddr is stored for callee context to avoid recomputation. Unfortunately it results in some bugs (eg. CopyRegDisplay doesn't transfer the information correctly, IsInCalleesFrames depends on quirks in the initial state, etc.) and I opted not to pursue this optimization at this time. I left the helper methods in place to make it easier to address this in future and to create a central place where we can _ASSERTE on invalid values.
…e and thus avoid repeated GetGCInfoToken() calls
|
/ba-g the test failure has been happening consistently on all runs since yesterday. #114426 |
On all other platforms CodeManState was unused and taking space in StackFrameIterator.
Optimize some code paths to cache GC info early or decode the method size without decoding the whole table.