[PATCH] D119792: [Clang] [P2025] Analyze only potential scopes for NRVO

[Chuanqi Xu via Phabricator via cfe-commits]

ChuanqiXu accepted this revision.
ChuanqiXu added inline comments.
This revision is now accepted and ready to land.


================
Comment at: clang/lib/Sema/Scope.cpp:152-154
+  // Consider the variable as NRVO candidate if the return slot is available
+  // for it in the current scope, or if it can be available in outer scopes.
+  NRVO = CanBePutInReturnSlot ? VD : nullptr;
----------------
rusyaev-roman wrote:
> ChuanqiXu wrote:
> > rusyaev-roman wrote:
> > > ChuanqiXu wrote:
> > > > What if NRVO contains a value already? It is possible due to the value of NRVO could be set by its children.
> > > Actually this is intention. If the parent has already NRVO candidate, then it should be invalidated (or not). Let's consider the following examples:
> > > 
> > > 
> > > ```
> > > X foo(bool b) {
> > >    X x;
> > >    X y;
> > >    if (b)
> > >       return x;
> > >    else
> > >       return y; // when we process this return statement, the parent has already NRVO and it will be invalidated (this is correct behavior)
> > > }
> > > ```
> > > 
> > > ```
> > > X foo(bool b) {
> > >    X x;
> > >    if (b)
> > >       return x;
> > >    
> > >    X y;
> > >    // when we process this return statement, the parent has already NRVO and it WON't be invalidated
> > >    //  (this is correct behavior), because a return slot will be available for it
> > >    return y;
> > > }
> > > ```
> > > 
> > > ```
> > > X foo(bool b) {
> > >    X x;
> > >    if (b)
> > >       return x;
> > > 
> > >    // when we process this return statement, the parent has already NRVO and it WON't be invalidated (this is correct behavior)
> > >    return x;
> > > }
> > > ```
> > > 
> > > ```
> > > X foo(bool b, X x) {
> > >    X y;
> > >    
> > >    if (b)
> > >       return x;
> > > 
> > >    // when we process this return statement, the parent contains nullptr (invalid candidate) and it will be invalidated (this is correct behavior)
> > >    return y;
> > > }
> > > ```
> > > 
> > > ```
> > > X foo(bool b, X x) {
> > >    if (b)
> > >       return x;
> > > 
> > >    X y;
> > >    // when we process this return statement, the parent contains nullptr (invalid candidate) and it WON't be invalidated (this is correct behavior)
> > >    return y;
> > > }
> > > ```
> > Oh, I see. Tricky. I don't find invalid cases now. But I recommend to comment that the children would maintain the `ReturnSlots` of their parents. (This is anti-intuition)
> > 
> > Have you tested any larger projects? Like libc++, libstdc++ or something like folly. I feel we need to do such tests to avoid we get anything wrong.
> I've already added a comment at the beginning of `updateNRVOCandidate` function where this point is mentioned: 
> ```
> //      ... Therefore, we need to clear return slots for other
> //      variables defined before the current return statement in the current
> //      scope and in outer scopes.
> ```
> If it's not enough, please let me know.
> 
> 
> > Have you tested any larger projects?
> 
> Yes, I've built the `clang` itself and `compiler-rt` project. Then I've checked them to run 'check-all' (on built clang and compiler-rt). Everything  works.
Great! Clang should be large enough.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D119792/new/

https://reviews.llvm.org/D119792