[LLVMdev] [WinEH] Cloning blocks that reference non-cloned PHI nodes

[Reid Kleckner]

Yeah, wrapping up the whole selector comparison into one intrinsic would
help. Thanks for the idea. :)

The current strategy also seems pretty fragile around complex cleanups. If
the cleanup isn't some simple destructor call with easily traceable
unconditional branch control flow, we get lost and can't find our way to
the next selector comparison. Having an explicit 'resume' or some other
intrinsic at the end of the cleanup would be pretty helpful here.

---

We also talked a lot about making WinEHPrepare try to less work all at
once, and split it into phases. This is separable from the landingswitch
proposal, which is more about making it easier to identify handler start
points. Here's a sketch of the phases we came up with:

Step 1: Identify selector dispatch conditional branches. Identify all
cleanup actions, and split basic blocks to make them start and end on a
basic block boundary. Build a list of all basic blocks that start a handler.

Step 2: Do a forwards-backwards dataflow analysis to mark basic blocks as
reachable and unreachable from each handler start block. The forwards part
is an easy DFS, the backwards part is basically about pruning unreachable
blocks like 'ret' from a handler. We can probably skip the backwards part
initially.

Step 3: Clone all blocks that are reachable from more than one handler
until all blocks belong to exactly one handler, and sink code from
landingpads back into handlers if necessary. We don't have many code
commoning optimizations, so this step is more of a defense against
theoretical future optimization passes that do more than tail merging.

Step 4: Demote all SSA values live from one handler into or out of another.
The only SSA values referenceable across handlers are static allocas.

Step 5: Extract the blocks belonging to each handler into their own
functions. Rewrite static alloca references in each new handler to use
llvm.framerecover.

Step 6: Profit?

Does this seem like a reasonable plan? IMO simplifying the overarching
design is a higher priority right now than fixing individual test cases.

On Fri, Apr 10, 2015 at 4:36 PM, Kaylor, Andrew <andrew.kaylor at intel.com>
wrote:

>  Just as an off-the-cuff idea that I haven’t thought through at all, what
> would you think of the possibility of replacing this:
>
>
>
>   %12 = call i32 @llvm.eh.typeid.for(i8* bitcast (%eh.CatchHandlerType*
> @llvm.eh.handlertype.H.0 to i8*))
>
>   %matches7 = icmp eq i32 %sel6, %12
>
>
>
> with this:
>
>
>
>   %matches7 = call i32 @llvm.eh.cmp.selector(i32 %sel6, i8* bitcast
> (%eh.CatchHandlerType* @llvm.eh.handlertype.H.0 to i8*))
>
>
>
> If we did that, the outlining code could just assume %sel6 was the current
> selector (because we don’t clone into nested landing pads) and we wouldn’t
> need to track it at all.
>
>
>
> -Andy
>
>
>
> *From:* Reid Kleckner [mailto:rnk at google.com]
> *Sent:* Friday, April 10, 2015 4:06 PM
> *To:* Kaylor, Andrew
> *Cc:* David Majnemer <david.majnemer at gmail.com> (david.majnemer at gmail.com);
> LLVM Developers Mailing List
> *Subject:* Re: [WinEH] Cloning blocks that reference non-cloned PHI nodes
>
>
>
> Sounds reasonable. Honestly, we've been having some serious whiteboard
> discussions about how to get out of doing all this selector pattern
> matching. I think that's basically the most fragile part of winehprepare,
> and if we can add some new constructs that wrap that up in something opaque
> to mid-level passes, that would be great.
>
>
>
> Unfortunately, I haven't really been able to get enough consensus to write
> something down. If I can't convince my coworkers that we need a new
> construct in person, there's not much point in writing up a detailed RFC. :(
>
>
>
> Here's the gist of the new idea:
>
> - Leave invoke alone, too many optimizations care about it
>
> - Add a new 'landingswitch' (name under debate) instruction that fills the
> landingpad role and is also a terminator
>
> - landingswitch clauses have some fixed operands (RTTI, handler label) and
> some variable number of operands as required by the EH personality (catch
> object, adjectives)
>
> - Allow the 'resume' instruction to have 0, 1, or 2 successors, one of
> which can be a landingpad.
>
> - Add a 'resume' clause to landingswitches to allow them to continue the
> unwind process at another landingpad in the same function
>
>
>
> I've tried to sell this idea, but I'm getting push back suggesting that we
> fix it a different way. The most compelling suggestion is to have the
> existing landingpad instruction produce an i8* selector instead of an i32
> selector, and then we indirectbr on it. This would save the selector
> pattern matching that we do today, but I don't particularly like it.
> Optimization passes can do basic block commoning, tail merging, and hoist
> instructions between the landingpad and the indirectbr, and I don't like
> dealing with that.
>
>
>
> I also don't like it because we'd need to teach the inliner about
> indirectbr, which today it refuses to inline.
>
>
>
> On Fri, Apr 10, 2015 at 3:45 PM, Kaylor, Andrew <andrew.kaylor at intel.com>
> wrote:
>
> Hi Reid and David,
>
>
>
> I just wanted to give you a heads up that I’m currently working on a
> problem where the WinEHPrepare outlining code stumbles and asserts while
> cloning code that references extracted landing pad values via PHI nodes
> that are in intermediate blocks that aren’t being cloned.  The test I’m
> looking at fails with an assertion claiming that llvm.eh.begincatch was
> called twice inside a handler.
>
>
>
> I have an idea for how to address this (giving the cloning director a
> chance to replace operands before an instruction is remapped).  I’ll put
> something up for review as soon as I have it working.  In the mean time I
> wanted to make sure you weren’t working on the same problem.
>
>
>
> I’m seeing it with some C++ EH tests I’m working on, but I think it may be
> possible for it to happen with SEH cleanup.
>
>
>
> If you’re curious, here’s a reproducer.
>
>
>
> int main(void) {
>
>   try {
>
>     try {
>
>       throw 'a';
>
>     } catch (char c) {
>
>       printf("%c\n", c);
>
>     }
>
>     throw 1;
>
>   } catch(int x) {
>
>     printf("%d\n", x);
>
>   } catch(...) {
>
>     printf("...\n");
>
>   }
>
>
>
>   try {
>
>     try {
>
>       throw 'b';
>
>     } catch (char c) {
>
>       printf("%c\n", c);
>
>     }
>
>     throw 2;
>
>   } catch(int x) {
>
>     printf("%d\n", x);
>
>   } catch (char c) {
>
>     printf("%c\n", c);
>
>   } catch(...) {
>
>     printf("...\n");
>
>   }
>
>   return 0;
>
> }
>
>
>
> -Andy
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150410/93a60489/attachment.html>