[PATCH] D130600: [clang][dataflow] Handle return statements

[Sam Estep via Phabricator via cfe-commits]

samestep added inline comments.


================
Comment at: clang/lib/Analysis/FlowSensitive/Transfer.cpp:348
+    auto *Loc = Env.getReturnStorageLocation();
+    assert(Loc != nullptr);
+    // FIXME: Model NRVO.
----------------
sgatev wrote:
> samestep wrote:
> > sgatev wrote:
> > > Let's do `if (Loc == nullptr) return;`
> > I don't think we want to do that, right? Shouldn't the `return` storage location always be set? Or is this about the "analyzing fragments rather than full functions" thing we discussed yesterday?
> I think it's related. If we are going with always initializing the `return` storage location then I guess at some point we should be able to make `Environment::getReturnStorageLocation` return a reference? In that case I'm fine with keeping the assert around in the meantime.
OK; yeah, I think the intention is that we're always initializing it. I'll leave this code as is for now, then.


================
Comment at: clang/lib/Analysis/FlowSensitive/Transfer.cpp:564
+      assert(ReturnLoc != nullptr);
+      Env.setStorageLocation(*S, *ReturnLoc);
+      Env.popCall(ExitEnv);
----------------
sgatev wrote:
> samestep wrote:
> > sgatev wrote:
> > > We use stable storage locations to ensure convergence. In that spirit, shouldn't we assign `ReturnLoc`'s value to `S`'s storage location instead of changing the storage location? Alternatively, we can pass `S`'s storage location to `pushCall` so that it can store it as `ReturnLoc`.
> > Could you clarify how this hurts convergence? My understanding is that `ReturnLoc` here is already stable, so this would make `S`'s storage location stable too.
> If I follow correctly, `ReturnLoc` here is the result of `Env.createStorageLocation(ReturnType)` which isn't stable. Each call to `createStorageLocation` returns a fresh storage location.
Ah I see, you're right. Is there a way to make a stable storage location for the `return`? My intuition is that we can't just pass `S`'s storage location to `pushCall`, because we want the storage location for the `return` to be the same across analysis of different callsites to the callee (similar to how we currently use the same storage location for a given parameter of the callee, regardless of how many times we analyze it). But maybe it would be fine; @ymandel do you have any thoughts on this?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130600/new/

https://reviews.llvm.org/D130600