https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126191
--- Comment #3 from migraineman33 at gmail dot com --- Thanks — that matches our reading: the defect is in the expander itself (ZERO_EXTEND where SIGN_EXTEND belongs), which is unchanged on trunk; r15-2196's ext-dce happens to keep this shape away from it, so trunk masks rather than fixes it, while the 13/14 release branches reproduce. The patch still corrects the backend, and it suggests the eventual testsuite case should pin itself against the mid-end transform (e.g. disable ext-dce, or scan the expander's output shape) so it keeps testing the backend rather than the optimizer's mood.
