JIT: Avoid changing def registers that may assume an implicit kill#127184
JIT: Avoid changing def registers that may assume an implicit kill#127184jakobbotsch merged 1 commit intodotnet:mainfrom
Conversation
LSRA build has places where it assumes that `BuildDef(tree, SRBM_REG)` means that `SRBM_REG` is going to be spilled if something is in it, without constructing any explicit kill. This assumption is incompatible with the logic in `resolveConflictingDefAndUse` that changes the def register on fixed reg definitions -- the required spilling will happen only if there ends up being a def into that register. We could fix all the places that assume this and add explicit kills, but instead this change just avoids changing the def assignment away from the register when it detects there is an active interval assigned to the fixed def register.
github-actions
Bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
There was a problem hiding this comment.
Pull request overview
Adjusts LSRA’s conflicting def/use resolution to avoid changing the assigned register of fixed-reg defs when doing so would prevent an implicit spill of a currently-occupied fixed register. This targets stress-induced miscompilations/test failures reported in #127130 and #127131.
Changes:
- Introduce a
canChangeDefgate to prevent def-reg reassignment for multireg nodes. - Further restrict def-reg reassignment for fixed-reg defs when the fixed register currently holds an active interval (ensuring the expected spill occurs).
|
cc @dotnet/jit-contrib PTAL @AndyAyersMS No diffs. Verified this fixes the two failures locally. Slowly learning LSRA intricacies one stress failure at a time... |
|
@dhartglassMSFT Maybe you can take a look at this? |
tannergooding
left a comment
tannergooding
left a comment
There was a problem hiding this comment.
This looks good/correct to me with what I understand of LSRA
LSRA build has places where it assumes that
BuildDef(tree, SRBM_REG)means thatSRBM_REGis going to be spilled if something is in it, without constructing any explicit kill. This assumption is incompatible with the logic inresolveConflictingDefAndUsethat changes the def register on fixed reg definitions -- the required spilling will happen only if there ends up being a def into that register.We could fix all the places that assume this and add explicit kills, but instead this change just avoids changing the def assignment away from the register when it detects there is an active interval assigned to the fixed def register.
Fix #127130
Fix #127131