Peter Kästle [off-list ref] writes:

On quick glance this sounds plausible, but to fully understand it I
need to put some effort in reading this code again.  I hope to do so 
tomorrow.  We can then compile a new set of patches including this
real fix and Ralf's and my test case.

Thanks for digging into it.

Yeah, thanks, both for noticing problem so quickly and started
digging to find the real solution.

By the way, as to the "if this were originally two patches, we could
have saved the test that expects failure" you raised, I do not think
it is a good idea to optimize for cases where changes turn out
problematic and have to get reverted.