avoid gcc optimizer bug by not force inlining part of `thrust::transf…

…orm` (#2509)
NVIDIA · Oct 4, 2024 · 583567b · 583567b
1 parent e8d57c3
commit 583567b
Showing 1 changed file with 6 additions and 2 deletions.
diff --git a/thrust/thrust/system/cuda/detail/transform.h b/thrust/thrust/system/cuda/detail/transform.h
@@ -178,8 +178,10 @@ struct binary_transform_f<InputIt1, InputIt2, OutputIt, no_stencil_tag, Transfor
   }
 }; // struct binary_transform_f
 
+// EAN 2024-10-04: when force-inlined, gcc's optimizer will generate bad code
+// for this function:
 template <class Policy, class InputIt, class Size, class OutputIt, class StencilIt, class TransformOp, class Predicate>
-OutputIt THRUST_FUNCTION unary(
+OutputIt _CCCL_HOST_DEVICE inline unary(
   Policy& policy,
   InputIt items,
   OutputIt result,
@@ -200,6 +202,8 @@ OutputIt THRUST_FUNCTION unary(
   return result + num_items;
 }
 
+// EAN 2024-10-04: when force-inlined, gcc's optimizer will generate bad code
+// for this function:
 template <class Policy,
           class InputIt1,
           class InputIt2,
@@ -208,7 +212,7 @@ template <class Policy,
           class StencilIt,
           class TransformOp,
           class Predicate>
-OutputIt THRUST_FUNCTION binary(
+OutputIt _CCCL_HOST_DEVICE inline binary(
   Policy& policy,
   InputIt1 items1,
   InputIt2 items2,