https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121661
--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> --- (In reply to Andrew Pinski from comment #7) > (In reply to Hongtao Liu from comment #6) > > Looks correct in the gimple > > The bug only happens at -O0. At higher levels it is ok. For my reduced testcase we see this after complex lowering: ``` _9 = REALPART_EXPR <FRAME.0.val>; _10 = IMAGPART_EXPR <FRAME.0.val>; _5 = COMPLEX_EXPR <_9, _10>; FRAME.0.FRAME_BASE.PARENT = _4; REALPART_EXPR <FRAME.0.val> = _9; IMAGPART_EXPR <FRAME.0.val> = _10; _1 = REALPART_EXPR <FRAME.0.val>; ``` It was originally: ``` _5 = val; FRAME.0.FRAME_BASE.PARENT = _4; FRAME.0.val = _5; _1 = REALPART_EXPR <FRAME.0.val>; ``` Looks like val is being replaced with FRAME.0.val which is incorrect. Yes I think there is a DECL_EXPR_DECL on val DECL. though it is NOT correct at this point. I think the problem is more related to use of force_gimple_operand_gsi here. Which is not needed for a simple DECL. extract_component can create the gimple assignment itself. This code dates long time ago which is why it is most likely using force_gimple_operand_gsi here.
