See PR36593. In one of the comments, I mention that the testcase in that PR (which was originally marked as a miscompilation) had warnings for -Wstrict-aliasing=1 and 2, but not 3. According to the documentation for -Wstrict-aliasing, level 3 ought to have *less* false negatives than 1 and 2, even more important when "bad code" was actually generated. Hence this enhanceement PR.
It could even be stretched to be seen as a regression: the warnings no longer match the generated code.