https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128868
--- Comment #7 from Torbjorn SVENSSON <azoff at gcc dot gnu.org> --- Following up on comment #6: I found the same operand-overlap problem in the NEON copysign<mode>3 expander in neon.md, for both V2SF and V4SF. When out-of-SSA coalescing makes the destination and magnitude operand share a pseudo, the initial copy of the sign operand overwrites the magnitude before it is read. The operation then returns the sign operand instead of the expected copysign result. Unlike the scalar issue, which affects soft float, the NEON issue affects softfp/hard. My earlier statement that softfp/hard are unaffected was therefore only correct for the scalar case. I've amended the commit in https://forge.sourceware.org/gcc/gcc/pulls/262 to address the NEON case too.
