On Fri, Aug 21, 2015 at 2:49 PM, Marek Polacek <[email protected]> wrote:
> On Fri, Aug 21, 2015 at 01:27:43PM +0200, Richard Biener wrote:
>> On Fri, Aug 21, 2015 at 12:52 PM, Marek Polacek <[email protected]> wrote:
>> > This fixes the libgo breakage. Seems I really should have removed the
>> > edge after we split the block with null dereference after __builtin_trap
>> > statement so that it's unreachable.
>> >
>> > Bootstrapped/regtested on x86_64-linux + ppc64-linux + bootstrapped on
>> > aarch64-linux, ok for trunk?
>>
>> Hum. I don't see why this is needed - CFG cleanup (which of course needs
>> to run!) should do this for you. In fact stray unreachable blocks are
>> usually
>> more of a problem.
>
> Aha. It seems cleanup does that if I change the code to generate
> __builtin_unreachable instead of __builtin_trap. A hint maybe? ;)
Not really...
static bool
cleanup_control_flow_bb (basic_block bb)
{
...
/* Check for indirect calls that have been turned into
noreturn calls. */
else if (is_gimple_call (stmt)
&& gimple_call_noreturn_p (stmt)
&& remove_fallthru_edge (bb->succs))
retval = true;
and __builtin_trap is NORETURN. But there is the hint where to debug.
Richard.
> What's going on here is that we find a potential NULL dereference
> (PHI argument has NULL), we call isolate_path, that duplicates a bb
> and then cuts off the outgoing edges of this duplicated bb. And in
> some cases it leaves COND_EXPR at the end of the bb -> ICE.
> The we insert __builtin_trap on this isolated path.
>
> Marek
