[Flang][OpenMP][Semantics] Modify errors to warnings for semantic checks in IS_DEVICE_PTR related to list-items being dummy arguments.

[raghavendhra]

This change is blocker for #71255

Changed semantic check from giving error to giving a warning about deprecation from OpenMP 5.2 and later about checks for dummy argument list-items present on IS_DEVICE_PTR clause.

[llvmbot]

@llvm/pr-subscribers-flang-semantics

@llvm/pr-subscribers-flang-openmp

Author: Raghu Maddhipatla (raghavendhra)

Changes

Modifying #71255

Changed semantic check to only check for, "Variable in IS_DEVICE_PTR clause must be a dummy argument that does not have the ALLOCATABLE, POINTER or VALUE attribute." Earlier there was an additional semantic check "Variable in IS_DEVICE_PTR clause must be a dummy argument" which might be incorrect.

---

Full diff: https://github.com/llvm/llvm-project/pull/74370.diff

2 Files Affected:

• (modified) flang/lib/Semantics/check-omp-structure.cpp (+2-5)
• (modified) flang/test/Semantics/OpenMP/target01.f90 (-1)

diff --git a/flang/lib/Semantics/check-omp-structure.cpp b/flang/lib/Semantics/check-omp-structure.cpp
index 2f4eb9a854270..c2362d225130d 100644
--- a/flang/lib/Semantics/check-omp-structure.cpp
+++ b/flang/lib/Semantics/check-omp-structure.cpp
@@ -2994,11 +2994,8 @@ void OmpStructureChecker::Enter(const parser::OmpClause::IsDevicePtr &x) {
         context_.Say(itr->second->source,
             "Variable '%s' in IS_DEVICE_PTR clause must be of type C_PTR"_err_en_US,
             source.ToString());
-      } else if (!(IsDummy(*symbol))) {
-        context_.Say(itr->second->source,
-            "Variable '%s' in IS_DEVICE_PTR clause must be a dummy argument"_err_en_US,
-            source.ToString());
-      } else if (IsAllocatableOrPointer(*symbol) || IsValue(*symbol)) {
+      } else if (IsDummy(*symbol) &&
+          (IsAllocatableOrPointer(*symbol) || IsValue(*symbol))) {
         context_.Say(itr->second->source,
             "Variable '%s' in IS_DEVICE_PTR clause must be a dummy argument that does not have the ALLOCATABLE, POINTER or VALUE attribute."_err_en_US,
             source.ToString());
diff --git a/flang/test/Semantics/OpenMP/target01.f90 b/flang/test/Semantics/OpenMP/target01.f90
index 485fa1f2530c3..2ce9a1af7cc80 100644
--- a/flang/test/Semantics/OpenMP/target01.f90
+++ b/flang/test/Semantics/OpenMP/target01.f90
@@ -39,7 +39,6 @@ subroutine bar(b1, b2, b3)
   type(c_ptr), pointer :: b2
   type(c_ptr), value :: b3
 
-  !ERROR: Variable 'c' in IS_DEVICE_PTR clause must be a dummy argument
   !$omp target is_device_ptr(c)
     y = y + 1
   !$omp end target

[kiranchandramohan]

Could you add a reference to the Section from the standard that asks for this restriction?

[raghavendhra]

Could you add a reference to the Section from the standard that asks for this restriction?

https://www.openmp.org/spec-html/5.0/openmpsu60.html

A list item that appears in an is_device_ptr clause must be a dummy argument that does not have the ALLOCATABLE, POINTER or VALUE attribute.

[shraiysh]

According to the restriction, this error should be reported right? It says argument must be a dummy argument. Am I missing something?

[raghavendhra]

I think this restriction is something on the lines of "If the list-item in an is_device_ptr clause is a dummy argument it should not have ALLOCATABLE, POINTER or VALUE attributes."

I think this following text might not have completely meant that all list-items must be dummy arguments only I think the conjunction of "does not have the ALLOCATABLE, POINTER or VALUE attribute" might be the reason for the "must".
A list item that appears in an is_device_ptr clause must be a dummy argument that does not have the ALLOCATABLE, POINTER or VALUE attribute.

Also, this restriction is not present from OpenMP 5.1 onwards through which I assumed it might not be originally intended for OpenMP spec committee might not have intended to restricted list-item of is_device_ptr clause to be dummy argument only.

Please correct me if I am interpreting it incorrectly. :)

[NimishMishra]

So it seems like the type of list item became more restrictive as the standard progressed.

OpenMP 4.5 standard (2.10.4): A list item that appears in an is_device_ptr clause must be a dummy argument

OpenMP 5.0 standard (2.12.5): A list item that appears in an is_device_ptr clause must be a dummy argument that does not have the ALLOCATABLE, POINTER or VALUE attribute.

However, my understanding of both these statements is leaning similar to @shraiysh's- any list item in this clause has to be a dummy argument, but without any additional attributes from ALLOCATABLE, POINTER, VALUE. In other words, I don't think non-dummy arguments are allowed in this clause. This becomes clear when we look at the OpenMP 4.5 standard wording: it is, without a doubt, restricting all list items within the clause to be dummy arguments.

[shraiysh]

Like Raghu mentioned, I cannot find this check, or anything related to the c_ptr being a dummy argument in the OpenMP 5.2 spec. If this check is not required in 5.2, we probably should remove this check completely - please let me know if there is anything related to this in 5.2. As discussed in the meeting, OpenMP 5.2 is essentially a fix on OpenMP 5.1 and OpenMP 5.0, so we should stick to OpenMP 5.2.

[raghavendhra]

I asked @mjklemm who is CEO of OpenMP ARB

Here are his comments:

@mjklemm : 5.0 is broken. 5.1/5.2 has fixed it. OpenMP 6.0 will have a different behavior. I wonder if it would make sense to directly implement the 5.2 and 6.0 behavior.

This aligns with @shraiysh 's comment. Even I could not find anything related to dummy argument list item of IS_DEVICE_PTR clause. So, I will remove the dummy argument check.

[kkwli]

I think in 5.2 we deprecate all the other use of the is_device_ptr clause in Fortran except type(c_ptr) variables in order to avoid any confusion between pointer and nonpointer list item. The has_device_address clause should be the choice for Fortran unless it is a type(c_ptr) variable.

In Section 5.4.7, we have:

If the is_device_ptr clause is specified on a target construct, if any list item is not of type C_PTR, the behavior is as if the list item appeared in a has_device_addr clause. Support for such list items in an is_device_ptr clause is deprecated.

In my opinion, issuing a warning message about the use being deprecated is good for the users. Furthermore, I don't think the spec wants to add more Fortran stuff to the is_device_ptr clause as it creates so much confusion in the past.

[raghavendhra]

I followed all your review comments and updated my patch. Can you please review? Thanks!

[kiranchandramohan]

LG.