[PATCH] D60716: [DwarfDebug] Dump call site debug info into DWARF

[Ulrich Weigand via Phabricator via llvm-commits]

uweigand added a comment.

This commit broke codegen on SystemZ, we're now failing both the test-suite and the multistage bootstrap.  Unfortunately this wasn't detected earlier since the SystemZ build bot was down.

The regressions go away when I revert the change to TargetRegisterInfo::isCallerPreservedPhysReg, i.e. change it back to always returning false.   Note that this change is the only modification in this patch that has any effect on actual code generation (as opposed to debug info generation) -- was this really intentional?

I have not yet debugged the failure in detail, but I believe the isCallerPreservedPhysReg change is in fact incorrect -- this function was not intended to just mirror getCallPreservedMask (which simply describes the function call ABI), but rather to handle special cases like the PowerPC TOC register, which is not just simply preserved across calls, but rather is invariant throughout the whole function.  This is why this check is used to eliminate invariant stores in MachineLICM.  The change in this patch breaks that semantics.

This really needs to be fixed before the LLVM 9 release.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D60716/new/

https://reviews.llvm.org/D60716