https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70593
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- And there is another thing special on this testcase - a_6 has zero uses, so isn't live. If both the values would be live after the asm, there would be an conflict added, even during the SSA_OP_DEF processing of one of the defs vs. the other one. So, I see generally two solutions: ensure conflict between multiple SSA_OP_DEFs of GIMPLE_ASM even if they aren't live, or during expansion for output operands of GIMPLE_ASM which have zero uses don't try to move them to any pseudo etc.
