On Mon, Apr 26, 2021 at 06:49:22PM +0200, Eric Botcazou wrote:
> the compiler can synthesize DWARF functions to describe the location and size
> of components in discriminated record types with variant part in Ada, but in
> peculiar cases the compiler drops expressions on the floor, for example in
> the divide case:
>
> case ROUND_DIV_EXPR:
> case TRUNC_DIV_EXPR:
> case EXACT_DIV_EXPR:
> if (TYPE_UNSIGNED (TREE_TYPE (loc)))
> return 0;
> op = DW_OP_div;
> goto do_binop;
>
> Now sizetype and bitsizetype are unsigned types, which means that any divide
> expression in them is dropped. But the sign bit is expected to be never set
> for sizetype or bitsizetype quantities so that's unnecessarily cautious.
>
> Tested on x86-64/Linux, both GCC and GDB, OK for the mainline?
At least sizetype is for GIMPLE compatible with size_t or unsigned long (or
whatever unsigned type has the same precision), so this looks wrong to me.
For non-strict DWARF or DWARF5 and above, I don't see why we can't use
typed DWARF ops instead though, see what e.g. mem_loc_descriptor does for
UDIV.
Jakub
