[Lldb-commits] [PATCH] D78077: Fix bug in UnwindAssemblyInstEmulation with fp-using codegen and mid-function epilogues

Jason Molenda via Phabricator via lldb-commits lldb-commits at lists.llvm.org
Mon Apr 13 21:10:27 PDT 2020

jasonmolenda created this revision.
jasonmolenda added reviewers: clayborg, labath.
jasonmolenda added a project: LLDB.
Herald added subscribers: danielkiss, kristof.beyls.

I found this on an AAarch64 target, where we use a frame pointer register on Darwin.  UnwindAssemblyInstEmulation iterates over the instructions building up an UnwindPlan row by row.  When UnwindAssemblyInstEmulation sees a branch instruction, it "forwards" the unwind state to the target offset; if we have a mid-function epilogue that restores the spilled registers and returns, when we hit the branch target, the saved unwind state is reinstated.

The bug comes in that UnwindAssemblyInstEmulation maintains both the "current row" object and it tracks the register that the Canonical Frame Address is defined in terms of, and whether this register is the frame pointer or not. After the prologue we have unwind state defining the CFA as $fp+16; then during a mid-function epilogue we restore the registers and switch to the CFA as $sp based.  At this point the UnwindAssemblyInstEmulation ivars (CFA register, is-it-fp-or-not) are updated to recognize the switch to $sp.  After the epilogue, we restore the current-row but we leave the other ivars to their epilogue setting.

The failure then comes in if the instructions after the mid-function epilogue modify $sp, UnwindAssemblyInstEmulation changes the current row's CFA offset value to track that change, because it thinks the CFA is defined in terms of the stack pointer.  But the current row is still defining it in terms of $fp.  So everything fails.

It's a sneaky bug because it's easy to miss if lldb silently "falls back" to the architectural default unwind plan when it can't find the caller.  So we're missing spilled registers from this frame.  It takes some specific circumstances to hit it -- modifying $sp after that mid-function epilogue.

I know not a lot of people have touched this code in ages (including me!), but let's put it up for review in case anyone has questions.

  rG LLVM Github Monorepo



-------------- next part --------------
A non-text attachment was scrubbed...
Name: D78077.257186.patch
Type: text/x-patch
Size: 8810 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/lldb-commits/attachments/20200414/86b01ee9/attachment.bin>

More information about the lldb-commits mailing list