Minimise modulos when rotating multidim arrays.

src/Futhark/CodeGen/ImpGen.hs

@@ -991,10 +991,11 @@ defCompileBasicOp _ Rearrange {} =
 defCompileBasicOp (Pat [pe]) (Rotate rs arr) = do
   shape <- arrayShape <$> lookupType arr
   sLoopNest shape $ \is -> do
-    is' <- sequence $ zipWith3 rotate (shapeDims shape) rs is
+    is' <- sequence $ zipWith3 rotate (shapeDims shape) (map pe64 rs) is
     copyDWIMFix (patElemName pe) is (Var arr) is'
   where
-    rotate d r i = dPrimVE "rot_i" $ rotateIndex (pe64 d) (pe64 r) i
+    rotate _ 0 i = pure i
+    rotate d r i = dPrimVE "rot_i" $ rotateIndex (pe64 d) r i
 defCompileBasicOp _ Reshape {} =
   pure ()
 defCompileBasicOp _ (UpdateAcc acc is vs) = sComment "UpdateAcc" $ do

src/Futhark/CodeGen/ImpGen/GPU/Base.hs

@@ -1418,6 +1418,7 @@ sRotateKernel dest rs src = do
       is'' <- sequence $ zipWith3 rotate ds rs is'
       copyDWIMFix dest is' (Var src) is''
   where
+    rotate _ 0 i = pure i
     rotate d r i = dPrimVE "rot_i" $ rotateIndex d r i
 
 compileThreadResult ::

src/Futhark/CodeGen/ImpGen/Multicore.hs

@@ -67,6 +67,7 @@ parallelRotate dest rs src = do
       is' <- dIndexSpace' "rep_i" ds i
       is'' <- sequence $ zipWith3 rotate ds rs is'
       copyDWIMFix dest is' (Var src) is''
+    rotate _ 0 i = pure i
     rotate d r i = dPrimVE "rot_i" $ rotateIndex d r i
 
 topLevelOps, inThreadOps :: Operations MCMem HostEnv Imp.Multicore