https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115426
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think this is a gimplification failure. 'r' is neither TREE_ADDRESSABLE
nor DECL_NOT_GIMPLE_REG and the =X constraint results in both allow_reg
and allow_mem but we gimplify it as is_gimple_lvalue which should,
as the base is a gimple register, emit a component extract to pre_p and
a complex build to post_p.
gimplify_compound_lval correctly sees this and forces a register argument
to the __imag operation but I'm not sure that's enough for lvalues.
IIRC a simple
__imag x = 1;
also doesn't have DECL_NOT_GIMPLE_REG on 'x', and gimplify_compound_lval
behaves the same. Still we eventually gimplify to
_1 = REALPART_EXPR <x>;
x = COMPLEX_EXPR <_1, 1.0e+0>;
D.2772 = x;
which is done via gimplify_modify_expr_complex_part. That suggests
it's gimplify_asm_expr that would need to do this very same thing as we
seem to rely on this for correctness.
With "=r" we correctly gimplify to
__asm__("" : "=r" D.2772);
_1 = REALPART_EXPR <r>;
r = COMPLEX_EXPR <_1, D.2772>;
D.2773 = r;
