On 11/19/19 5:03 PM, Jakub Jelinek wrote:
> On Tue, Sep 03, 2019 at 02:27:29PM -0600, Jeff Law wrote:
>>>> + /* The transformation below will inherently introduce a memory load,
>>>> + for which LHS may not be initialized yet if it is not in NOTRAP,
>>>> + so a -Wmaybe-uninitialized warning message could be triggered.
>>>> + Since it's a bit hard to have a very accurate uninitialization
>>>> + analysis to memory reference, we disable the warning here to avoid
>>>> + confusion. */
>>>> + TREE_NO_WARNING (lhs) = 1;
>>>
>>> I don't like this, but not for the reasons Martin stated, we use
>>> TREE_NO_WARNING not just when we've emitted warnings, but in many places
>>> when we've done something that might trigger false positives.
>>> Yes, it would be nice to do it more selectively.
>>>
>>> The problem I see with the above though is that lhs might very well be
>>> a decl, and setting TREE_NO_WARNING on it then doesn't affect only the
>>> hoisted load, but also all other code that refers to the decl.
>> LHS is restricted to just MEM_REF, ARRAY_REF and COMPONENT_REF. We'd be
>> setting the NO_WARNING bits on the toplevel expression, but not on
>> anything shared like a _DECL node.
>>
>> So what we're losing here would be things like out of bounds array
>> checks on the LHS, so it still sucks.
>
> Sorry for dropping the ball on this.
> You're right, LHS is a MEM_REF, ARRAY_REF or COMPONENT_REF, so what I was
> worried about doesn't happen.
> I've tried using gimple_set_no_warning, but tree-ssa-uninit.c in this case
> actually doesn't look at that (it does only in a different spot).
>
>>> If the TREE_NO_WARNING bit is set on something that isn't shareable, that is
>>> fine with me, like a MEM_REF, TARGET_MEM_REF or handled component. If lhs
>>> is a decl, can we force a MEM_REF around it (and won't we fold it back to
>>> the decl?)? Or perhaps better, can we gimple_set_no_warning on the load
>>> stmt instead?
>> We have the toplevel statement, so that might be worth a try as well.
>
> But, what the patch did was set it on the tree that is later unshared,
> which means TREE_NO_WARNING wasn't set just on the rhs1 of the load, but
> also on the lhs of the store.
>
> The following version fixes that and I've also added the testcase to the
> testsuite.
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2019-11-19 Jiangning Liu <jiangning....@amperecomputing.com>
> Jakub Jelinek <ja...@redhat.com>
>
> PR middle-end/91195
> * tree-ssa-phiopt.c (cond_store_replacement): Move lhs unsharing
> earlier. Set TREE_NO_WARNING on the rhs1 of the artificially added
> load.
>
> * gcc.dg/pr91195.c: New test.
OK
jeff
