Let rrev take and return dual components but not confuse cotangents

src/HordeAd/Core/TensorADVal.hs

@@ -475,10 +475,10 @@ instance ( Dual ranked ~ DeltaR ranked shaped
   ddummy = undefined
   dshape = undefined
 
-instance ( ADReadySmall (ADVal ranked) (ADVal shaped)
-         , DomainsTensor ranked shaped
-         , DualPart ranked
-         , Dual (Clown (DynamicOf ranked)) ~ DeltaD ranked shaped )
+instance ( ADReadyBoth (ADVal (ADVal ranked)) (ADVal (ADVal shaped))
+         , DualPart (ADVal ranked)
+         , Dual (Clown (DynamicOf (ADVal ranked)))
+           ~ DeltaD (ADVal ranked) (ADVal shaped) )
          => DomainsTensor (ADVal ranked) (ADVal shaped) where
   dmkDomains = id
   rletDomainsOf = (&)
@@ -489,8 +489,7 @@ instance ( ADReadySmall (ADVal ranked) (ADVal shaped)
   rrev :: (GoodScalar r, KnownNat n)
        => (forall f. ADReady f => Domains (DynamicOf f) -> f r n)
        -> DomainsOD
-       -> Domains (DynamicOf ranked)
-       -> DomainsOf ranked
+       -> Domains (DynamicOf (ADVal ranked))
+       -> DomainsOf (ADVal ranked)
   rrev f _parameters0 parameters =
-    dmkDomains @ranked $ fst
-    $ crevOnDomains Nothing (f @(ADVal ranked)) parameters
+    fst $ crevOnDomains Nothing (f @(ADVal (ADVal ranked))) parameters

src/HordeAd/Core/TensorAst.hs

@@ -412,8 +412,8 @@ instance AstSpan s => DomainsTensor (AstRanked s) (AstShaped s) where
   rrev :: GoodScalar r
        => (forall f. ADReady f => Domains (DynamicOf f) -> f r n)
        -> DomainsOD
-       -> Domains (AstDynamic PrimalSpan)
-       -> AstDomains PrimalSpan
+       -> Domains (AstDynamic s)
+       -> AstDomains s
   rrev f parameters0 domains =
     AstRev (funToAstDomains @PrimalSpan f parameters0)
            parameters0 (AstDomains domains)

src/HordeAd/Core/TensorClass.hs

@@ -600,13 +600,13 @@ class DomainsTensor (ranked :: RankedTensorKind)
   -- and the third has to have the same shapes as the second.
   --
   -- The function argument needs to be quantified (or an AST term),
-  -- because otherwise in the ADVal instance one could put illegal InputR there.
-  -- For the same reason there is PrimalOf in the last argument.
+  -- because otherwise in the ADVal instance one could put an illegal
+  -- InputR there, confusing two levels of contangents.
   rrev :: (GoodScalar r, KnownNat n)
        => (forall f. ADReady f => Domains (DynamicOf f) -> f r n)
        -> DomainsOD
-       -> Domains (DynamicOf (PrimalOf ranked))
-       -> DomainsOf (PrimalOf ranked)
+       -> Domains (DynamicOf ranked)
+       -> DomainsOf ranked
 
 
 -- * The giga-constraint
@@ -656,7 +656,11 @@ type ADReadySmall ranked shaped =
 
 type ADReadyBoth ranked shaped =
   ( ADReadySmall ranked shaped
-  , DomainsTensor ranked shaped
+-- TODO: this doesn't type-check and not because rrev uses it,
+-- but probably because ADVal instance of DomainsTensor uses ADReady
+-- at one more ADVal nesting level:
+--, DomainsTensor ranked shaped
+-- so we can't nest rrev right now
   , DomainsTensor (PrimalOf ranked) (PrimalOf shaped) )