[flang-commits] [flang] [flang][acc] Allow nested gang loops inside acc routines (PR #158693)

[via flang-commits]

https://github.com/khaki3 updated https://github.com/llvm/llvm-project/pull/158693

>From ce6bcb71669681c26c8c312ac4a1d0ffc71d6910 Mon Sep 17 00:00:00 2001
From: Kazuaki Matsumura <kmatsumura at nvidia.com>
Date: Mon, 15 Sep 2025 10:25:08 -0700
Subject: [PATCH 1/2] Allow nested gang clauses in acc routines

---
 flang/lib/Semantics/check-acc-structure.cpp |  8 +++---
 flang/lib/Semantics/check-acc-structure.h   |  2 +-
 flang/test/Semantics/OpenACC/acc-loop.f90   | 31 +++++++++++++++++++++
 3 files changed, 36 insertions(+), 5 deletions(-)

diff --git a/flang/lib/Semantics/check-acc-structure.cpp b/flang/lib/Semantics/check-acc-structure.cpp
index 6cb7e5e9e6e25..bcb485d4e0f47 100644
--- a/flang/lib/Semantics/check-acc-structure.cpp
+++ b/flang/lib/Semantics/check-acc-structure.cpp
@@ -136,10 +136,10 @@ void AccStructureChecker::CheckNotInComputeConstruct() {
   }
 }
 
-bool AccStructureChecker::IsInsideParallelConstruct() const {
+bool AccStructureChecker::IsInsideKernelsConstruct() const {
   if (auto directive = getParentComputeConstruct())
-    if (*directive == llvm::acc::ACCD_parallel ||
-        *directive == llvm::acc::ACCD_parallel_loop)
+    if (*directive == llvm::acc::ACCD_kernels ||
+        *directive == llvm::acc::ACCD_kernels_loop)
       return true;
   return false;
 }
@@ -293,7 +293,7 @@ void AccStructureChecker::CheckNotInSameOrSubLevelLoopConstruct() {
           bool invalid{false};
           if (parentClause == llvm::acc::Clause::ACCC_gang &&
               cl == llvm::acc::Clause::ACCC_gang) {
-            if (IsInsideParallelConstruct()) {
+            if (!IsInsideKernelsConstruct()) {
               auto parentDim = getGangDimensionSize(parent);
               auto currentDim = getGangDimensionSize(GetContext());
               std::int64_t parentDimNum = 1, currentDimNum = 1;
diff --git a/flang/lib/Semantics/check-acc-structure.h b/flang/lib/Semantics/check-acc-structure.h
index 711d0326349a4..09399297ca4be 100644
--- a/flang/lib/Semantics/check-acc-structure.h
+++ b/flang/lib/Semantics/check-acc-structure.h
@@ -101,7 +101,7 @@ class AccStructureChecker
   bool IsLoopConstruct(llvm::acc::Directive directive) const;
   std::optional<llvm::acc::Directive> getParentComputeConstruct() const;
   bool IsInsideComputeConstruct() const;
-  bool IsInsideParallelConstruct() const;
+  bool IsInsideKernelsConstruct() const;
   void CheckNotInComputeConstruct();
   std::optional<std::int64_t> getGangDimensionSize(
       DirectiveContext &dirContext);
diff --git a/flang/test/Semantics/OpenACC/acc-loop.f90 b/flang/test/Semantics/OpenACC/acc-loop.f90
index 77c427e0a85ae..d3777f48430e5 100644
--- a/flang/test/Semantics/OpenACC/acc-loop.f90
+++ b/flang/test/Semantics/OpenACC/acc-loop.f90
@@ -447,4 +447,35 @@ program openacc_loop_validity
     END DO
   END DO
 
+contains
+
+  subroutine sub1()
+    !$acc routine gang(dim:2)
+    implicit none
+    integer, parameter :: N = 256
+    integer :: i, j
+
+    !$acc loop gang(dim:2)
+    DO j = 1, N
+       !$acc loop gang(dim:1) vector
+       DO i = 1, N
+       END DO
+    END DO
+  end subroutine sub1
+
+  subroutine sub2()
+    !$acc routine gang(dim:2)
+    implicit none
+    integer, parameter :: N = 256
+    integer :: i, j
+
+    !$acc loop gang(dim:2)
+    DO j = 1, N
+       !ERROR: GANG(dim:2) clause is not allowed in the region of a loop with the GANG(dim:2) clause
+       !$acc loop gang(dim:2) vector
+       DO i = 1, N
+       END DO
+    END DO
+  end subroutine sub2
+
 end program openacc_loop_validity

>From 552e99faa483b7e8e125880470d6bd0cd2f10173 Mon Sep 17 00:00:00 2001
From: Kazuaki Matsumura <kmatsumura at nvidia.com>
Date: Tue, 16 Sep 2025 04:54:22 -0700
Subject: [PATCH 2/2] Reverse logic; Update the error message for the nested
 gang loops

---
 flang/lib/Semantics/check-acc-structure.cpp | 7 ++++---
 flang/test/Semantics/OpenACC/acc-loop.f90   | 4 ++--
 2 files changed, 6 insertions(+), 5 deletions(-)

diff --git a/flang/lib/Semantics/check-acc-structure.cpp b/flang/lib/Semantics/check-acc-structure.cpp
index bcb485d4e0f47..3cd6d6ba7689a 100644
--- a/flang/lib/Semantics/check-acc-structure.cpp
+++ b/flang/lib/Semantics/check-acc-structure.cpp
@@ -293,7 +293,10 @@ void AccStructureChecker::CheckNotInSameOrSubLevelLoopConstruct() {
           bool invalid{false};
           if (parentClause == llvm::acc::Clause::ACCC_gang &&
               cl == llvm::acc::Clause::ACCC_gang) {
-            if (!IsInsideKernelsConstruct()) {
+            if (IsInsideKernelsConstruct()) {
+              context_.Say(GetContext().clauseSource,
+                  "Nested GANG loops are not allowed in the region of a KERNELS construct"_err_en_US);
+            } else {
               auto parentDim = getGangDimensionSize(parent);
               auto currentDim = getGangDimensionSize(GetContext());
               std::int64_t parentDimNum = 1, currentDimNum = 1;
@@ -317,8 +320,6 @@ void AccStructureChecker::CheckNotInSameOrSubLevelLoopConstruct() {
                     parentDimStr);
                 continue;
               }
-            } else {
-              invalid = true;
             }
           } else if (parentClause == llvm::acc::Clause::ACCC_worker &&
               (cl == llvm::acc::Clause::ACCC_gang ||
diff --git a/flang/test/Semantics/OpenACC/acc-loop.f90 b/flang/test/Semantics/OpenACC/acc-loop.f90
index d3777f48430e5..635dbb04cd666 100644
--- a/flang/test/Semantics/OpenACC/acc-loop.f90
+++ b/flang/test/Semantics/OpenACC/acc-loop.f90
@@ -340,10 +340,10 @@ program openacc_loop_validity
 
   !$acc kernels loop gang(dim:3)
   do i = 1, n
-    !ERROR: GANG clause is not allowed in the region of a loop with the GANG clause
+    !ERROR: Nested GANG loops are not allowed in the region of a KERNELS construct
     !$acc loop gang(dim:2)
     do j = 1, n
-      !ERROR: GANG clause is not allowed in the region of a loop with the GANG clause
+      !ERROR: Nested GANG loops are not allowed in the region of a KERNELS construct
       !$acc loop gang(dim:1) worker vector
       do k = 1, i
       end do