[PATCH] D127487: [Sema] Fix assertion failure when instantiating requires expression

[Ilya Biryukov via Phabricator via cfe-commits]

ilya-biryukov added inline comments.


================
Comment at: clang/lib/Sema/SemaConcept.cpp:352
+      [this](const Expr *AtomicExpr) -> ExprResult {
+        // We only do this to immitate lvalue-to-rvalue conversion.
+        return PerformContextuallyConvertToBool(const_cast<Expr*>(AtomicExpr));
----------------
erichkeane wrote:
> ilya-biryukov wrote:
> > royjacobson wrote:
> > > erichkeane wrote:
> > > > ilya-biryukov wrote:
> > > > > erichkeane wrote:
> > > > > > Can you explain this more?  How does this work, and why don't we do that directly instead?
> > > > > That's entangled with `calculateConstraintSatisfaction`. I actually tried to do it directly, but before passing expressions to this function `calculateConstraintSatisfaction` calls `IgnoreParenImpCasts()`, which strips away the lvalue-to-rvalue conversion.
> > > > > And we need this conversion so that the evaluation that runs after this callback returns actually produces an r-value.
> > > > > 
> > > > > Note that the other call to `calculateConstraintSatisfaction` also calls `PerformContextuallyConvertToBool` after doing template substitution into the constraint expression.
> > > > > 
> > > > > I don't have full context on why it's the way it is, maybe there is a more fundamental change that helps with both cases.
> > > > Hmm... my understanding is we DO need these to be a boolean expression eventually, since we have to test them as a bool, so that is why the other attempts the converesion.  If you think of any generalizations of this, it would be appreciated, I'll think it through as well.
> > > Note we already have a related bug about this https://github.com/llvm/llvm-project/issues/54524
> > Yeah, they have to be bool and we actually check for that in `CheckConstraintExpression`. The standard explicitly mentions only the lvalue-to-rvalue conversion should be performed.
> > ```
> > [temp.constr.atomic]p3 If substitution results in an invalid type or expression, the constraint is
> > not satisfied. Otherwise, the lvalue-to-rvalue conversion (7.3.1) is performed if necessary, and E shall be a constant expression of type bool.
> > ```
> > 
> > However, in the calls to `calculateConstraintSatisfaction` we do a more generic boolean conversion, but the comment in the other call site suggests this probably accidental and we actually want a less generic conversion:
> > ```
> >           // Substitution might have stripped off a contextual conversion to
> >           // bool if this is the operand of an '&&' or '||'. For example, we
> >           // might lose an lvalue-to-rvalue conversion here. If so, put it back
> >           // before we try to evaluate.
> >           if (!SubstitutedExpression.isInvalid())
> >             SubstitutedExpression =
> >                 S.PerformContextuallyConvertToBool(SubstitutedExpression.get());
> > ```
> > 
> > I am happy to take a look at fixing the mentioned bug.
> Hmm... yeah, I find myself wondering if there is a better lval/rval conversion function here, and I'm guessing the contextually convert to bool is the wrong one.
Yep, it's definitely the wrong function. I could not find a better one, but I only briefly looked through the options.

I will have to find one or implement it while fixing  https://github.com/llvm/llvm-project/issues/54524.


================
Comment at: clang/lib/Sema/SemaTemplateInstantiate.cpp:2042
+        !SemaRef.CheckConstraintExpression(TransConstraint.get())) {
+      assert(Trap.hasErrorOccurred() && "CheckConstraintExpression failed, but "
+                                        "did not produce a SFINAE error");
----------------
erichkeane wrote:
> ilya-biryukov wrote:
> > erichkeane wrote:
> > > ilya-biryukov wrote:
> > > > erichkeane wrote:
> > > > > This branch ends up being empty if asserts are off.  Also, it results in CheckConstraintExpression happening 2x, which ends up being more expensive after https://reviews.llvm.org/D126907
> > > > > This branch ends up being empty if asserts are off.  Also, it results in CheckConstraintExpression happening 2x, which ends up being more expensive after https://reviews.llvm.org/D126907
> > > > 
> > > > Yeah, good point, I have update it.
> > > > 
> > > > I am not sure why would `CheckConstraintExpression` be called twice, could you elaborate? Note that we do not call `BuildNestedRequirement` anymore and use placement new directly to avoid extra template instantiations. Instead we call `CheckConstraintExpression` directly to account for any errors.
> > > This check does not seem to cause a 'return' to the function, but then falls through to the version on 2052, right?  
> > > 
> > > `CheckConstraintExpression`/`CheckConstraintSatisfaction`(i think?) ends up now having to re-instantiate every time it is called, so any ability to cache results ends up being beneficial here.
> > The number of calls to these functions is actually the same.
> > 
> > `CheckConstraintExpression` used to be called during `CheckConstraintSatisfaction` (that does instantiations) for every atomic constraint after the substitution. It merely checks that each constraint have a bool type and does not do any substitutions, so it's pretty cheap anyway.
> > 
> > `CheckConstraintSatisfaction` was called inside `BuildNestedRequirement`, we now call a different overload here directly that does not do any extra template instantiations directly.
> > 
> > That way we end up doing the same checks without running recursive template instantiations.
> Hmm... I guess what I'm saying is: I would like it if we could minimize/reduce the calls to calcuateConstraintSatisfaction (in SemaConcept.cpp) that does the call to SubstConstraintExpr (or substExpr).  
> 
> That ends up being more expensive thanks to my other patch.
Ah, yes, that makes sense. Note that `CheckConstraintSatisfaction` call on 2052 does not do any substitutions, it merely evaluates the expressions if they are not dependent.

I will have to look more closely into your patch to get a sense of why substituting to the constraint expressions is more expensive after it.

BTW, I was wondering if there is any reason to substitute empty template arguments to evaluate non-dependent constraints? (This is what `CalculateConstraintSatisfaction` does)
I feels like merely evaluating those should be enough.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D127487/new/

https://reviews.llvm.org/D127487