Stop tracking genReturnBB after global morph
#97130
Conversation
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThe After global morph, clear the pointer and don't maintain it; it is no This leads to diffs, as unreachable return blocks get deleted. This can
|
There was a problem hiding this comment.
LGTM, but seems like win-x86 is unhappy about the removal of (even unreachable) monitor exits in synchronized methods? Presumably in those cases we can just report the entire method (assuming the info is used to exit the lock on exceptional flow by the VM)?
BruceForstall
force-pushed
the
StopTrackingGenReturnBBAfterMorph
branch
from
26427c7 to
04f5bcb
Compare
|
win-x86 handling of synchronized methods is unique: they aren't transformed to try/finally, but instead are reported to the VM via the GC information and the VM handles unlocking. I'll look into this... |
|
/azp run runtime-coreclr outerloop, runtime-coreclr jitstress |
|
Azure Pipelines successfully started running 2 pipeline(s). |
The `genReturnBB` pointer is set if a merged return block is created during `fgAddInternal` (so, quite early during compilation). It needs to be maintained (and the unreferenced block it points to needs to be kept around) until global morph, which hooks up flow to this block. After global morph, clear the pointer and don't maintain it; it is no longer needed. Later code that was checking for it can instead check for `BBJ_RETURN` blocks or `GT_RETURN` nodes. Also, remove the marking of the `genReturnBB` block as `BBF_DONT_REMOVE`. This leads to diffs, as unreachable return blocks get deleted. This can happen, say, if all other exits are tail calls. If that happens and we're in a case of needing a profiler hook, then the profile exit hook is also not needed (it would be unreachable).
The JIT still needs to report an end-of-synchronized-range offset in the GC info for x86 synchronized methods, so just create a new label at the end of all blocks. The VM will automatically exit the synchronization on an exceptional method exit.
BruceForstall
force-pushed
the
StopTrackingGenReturnBBAfterMorph
branch
from
996b5b6 to
fca4c50
Compare
|
/azp run runtime-coreclr outerloop, runtime-coreclr jitstress |
|
Azure Pipelines successfully started running 2 pipeline(s). |
|
Pretty significant size improvements. |
* Stop tracking `genReturnBB` after global morph The `genReturnBB` pointer is set if a merged return block is created during `fgAddInternal` (so, quite early during compilation). It needs to be maintained (and the unreferenced block it points to needs to be kept around) until global morph, which hooks up flow to this block. After global morph, clear the pointer and don't maintain it; it is no longer needed. Later code that was checking for it can instead check for `BBJ_RETURN` blocks or `GT_RETURN` nodes. Also, remove the marking of the `genReturnBB` block as `BBF_DONT_REMOVE`. This leads to diffs, as unreachable return blocks get deleted. This can happen, say, if all other exits are tail calls. If that happens and we're in a case of needing a profiler hook, then the profile exit hook is also not needed (it would be unreachable). * Allow x86 synchronized methods to have unreachable return blocks The JIT still needs to report an end-of-synchronized-range offset in the GC info for x86 synchronized methods, so just create a new label at the end of all blocks. The VM will automatically exit the synchronization on an exceptional method exit. * Create final IG before clearing GC info
2 participants
The
genReturnBBpointer is set if a merged return block is createdduring
fgAddInternal(so, quite early during compilation). It needsto be maintained (and the unreferenced block it points to needs to be
kept around) until global morph, which hooks up flow to this block.
After global morph, clear the pointer and don't maintain it; it is no
longer needed. Later code that was checking for it can instead check
for
BBJ_RETURNblocks orGT_RETURNnodes. Also, remove the markingof the
genReturnBBblock asBBF_DONT_REMOVE.This leads to diffs, as unreachable return blocks get deleted. This can
happen, say, if all other exits are tail calls. If that happens and we're
in a case of needing a profiler hook, then the profile exit hook is also
not needed (it would be unreachable).