Since you only touch two tests in v4, I used the binaries that I built using v2 and tested the two test cases for my targets and they all pass. At least for the arm-none-eabi targets that I do test, I see no regressions with v4, so for me, it's a go.
As I do not know much about the .md magic, someone else needs to review the implementation. -- https://forge.sourceware.org/gcc/gcc/pulls/208#issuecomment-7056
