[PATCH] Proposal on how to fix temporary dtors.

[Manuel Klimek]

On Fri, Jun 20, 2014 at 5:28 AM, Jordan Rose <jordan_rose at apple.com> wrote:

>
> On Jun 19, 2014, at 14:19 , Manuel Klimek <klimek at google.com> wrote:
>
> On Thu, Jun 19, 2014 at 6:30 PM, Jordan Rose <jordan_rose at apple.com>
> wrote:
>
>> I think the algorithm makes sense. I'm not sure it's different, though,
>> than just passing up the first (or last) CXXBindTemporaryExpr visited for a
>> given expression, which would look like this:
>>
>>     // For a logical expression...
>>     VisitForTemporaryDtors(E->getLHS(), false, &LastBTE);
>>     const CXXBindTemporaryExpr *InnerBTE = nullptr;
>>     VisitForTemporaryDtors(E->getRHS(), false, &InnerBTE);
>>     InsertTempDtorDecisionBlock(InnerBTE);
>>
>> Are there any cases that wouldn't cover?
>>
>
> Well, the problem is what to do if we don't have a BTE yet; if we only
> wanted to pass one pointer, the semantics on function exit would need to be:
> if (BTE) { we have already found a BTE, no need to insert a new block when
> we encounter the next }
> else { we have not yet found a BTE, so we need to insert a new block when
> we find one }
>
> The "unconditional" branch would only fit with the first part, so we would
> need to somehow conjure non-null BTE for all unconditional entries, and
> then afterwards know that this is a special value, because since we didn't
> add an extra block (as we were running an unconditional case), we don't
> need a decision block.
> I'd say that's a pretty strong argument that we at least need to pass the
> CXXBindTemporaryExpr* and a bool IsConditional.
>
> Now it's right that we don't need to remember the Succ when we hit the
> conditional, and instead we could just always store the Succ when we enter
> a recursive visitation for a conditional branch (we need to store the Succ
> because we can have nested conditionals), but that seems to me like it
> would distribute the logic through even more places, and thus undesirable.
>
>
> My observation is that only certain expressions cause conditional
> branching, and that for these expressions you *always* need to introduce
> a new block if you find any destructors, say, in the RHS of a logical
> expression. So
>
> 1. if you're in a non-conditional sub-expression, it doesn't matter
> whether there are temporaries or not; you don't need to insert a decision
> branch,
> 2. if you're in a conditional sub-expression and there are no temporaries,
> you don't need to insert a decision branch,
> 3. if you're in a conditional sub-expression and there are temporaries,
> you can use any temporary from that subexpression as the guard.
>

That is exactly the algorithm.


> So it looks to me like the only information you have to pass up from
> traversing the sub-expressions is a BTE from that subexpression. Everything
> else can be handled at the point where you start processing that
> subexpression.
>

We have to pass the information down whether we are in a conditional part,
so we know whether we have to start a new block when we hit the temporary.

If you're asking why we cannot start the block at the conditional point,
the reason is that we cannot add it before we do the subexpression
traversal (because we don't know yet whether there will be temporaries, and
we don't want to add a block if there are no temporaries), and if we want
to do it after the subexpression traversal, we'd somehow need to split
blocks (as there can be nested conditionals, and multiple temporaries).

Cheers,
/Manuel


>
> What am I missing?
> Jordan
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140620/3a0abcc0/attachment.html>