[llvm-branch-commits] [flang] [flang][CUDA] Emit ignore_denormal_mode for CUDA Fortran atomicadd (PR #232726)

[Christian Sigg via llvm-branch-commits]

https://github.com/chsigg created https://github.com/llvm/llvm-project/pull/232726

None

>From f382e483219fef3194bd58633b1ed3567e194e16 Mon Sep 17 00:00:00 2001
From: Christian Sigg <csigg at google.com>
Date: Fri, 16 Oct 2026 08:16:41 +0200
Subject: [PATCH] [flang][CUDA] Emit ignore_denormal_mode for CUDA Fortran
 atomicadd

---
 .../Optimizer/Builder/CUDAIntrinsicCall.cpp   |  7 +++--
 flang/test/Lower/CUDA/cuda-atomicadd.cuf      | 30 +++++++++++++++++--
 flang/test/Lower/CUDA/cuda-device-proc.cuf    |  6 ++--
 3 files changed, 35 insertions(+), 8 deletions(-)

diff --git a/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp b/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
index e979bbc44af98f..4e61335079711e 100644
--- a/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
@@ -687,8 +687,11 @@ static mlir::Value genAtomBinOp(fir::FirOpBuilder &builder, mlir::Location &loc,
                                 mlir::Value arg1) {
   auto llvmPointerType = mlir::LLVM::LLVMPointerType::get(builder.getContext());
   arg0 = builder.createConvert(loc, llvmPointerType, arg0);
-  return mlir::LLVM::AtomicRMWOp::create(builder, loc, binOp, arg0, arg1,
-                                         mlir::LLVM::AtomicOrdering::seq_cst);
+  bool ignoreDenormalMode = binOp == mlir::LLVM::AtomicBinOp::fadd;
+  return mlir::LLVM::AtomicRMWOp::create(
+      builder, loc, binOp, arg0, arg1, mlir::LLVM::AtomicOrdering::seq_cst,
+      /*syncscope=*/llvm::StringRef(), /*alignment=*/0, /*isVolatile=*/false,
+      /*ignoreDenormalMode=*/ignoreDenormalMode);
 }
 
 // ATOMICADD
diff --git a/flang/test/Lower/CUDA/cuda-atomicadd.cuf b/flang/test/Lower/CUDA/cuda-atomicadd.cuf
index dff029d33930d4..39e5a56808a707 100644
--- a/flang/test/Lower/CUDA/cuda-atomicadd.cuf
+++ b/flang/test/Lower/CUDA/cuda-atomicadd.cuf
@@ -8,7 +8,7 @@ attributes(global) subroutine test_atomicaddvector_r2()
 end subroutine
 
 ! CHECK-LABEL: func.func @_QPtest_atomicaddvector_r2() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
-! CHECK: llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf16>
+! CHECK: llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf16>
 
 attributes(global) subroutine test_atomicaddvector_r4()
   real(4), device :: a(2), tmp1(2), tmp2(2)
@@ -16,7 +16,7 @@ attributes(global) subroutine test_atomicaddvector_r4()
 end subroutine
 
 ! CHECK-LABEL: func.func @_QPtest_atomicaddvector_r4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
-! CHECK: llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf32>
+! CHECK: llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf32>
 
 attributes(global) subroutine test_atomicadd_r2x4()
   real(4), device :: a(2), tmp1(2), tmp2(2)
@@ -24,7 +24,7 @@ attributes(global) subroutine test_atomicadd_r2x4()
 end subroutine
 
 ! CHECK-LABEL: func.func @_QPtest_atomicadd_r2x4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
-! CHECK: llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf32>
+! CHECK: llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf32>
 
 attributes(global) subroutine test_atomicadd_r4x4()
   real(4), device :: a(4), tmp1(4), tmp2(4)
@@ -33,3 +33,27 @@ end subroutine
 
 ! CHECK-LABEL: func.func @_QPtest_atomicadd_r4x4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
 ! CHECK: atom.add.v4.f32
+
+attributes(global) subroutine test_atomicadd_r4()
+  real(4), device :: a, tmp1, tmp2
+  tmp1 = atomicadd(a, tmp2)
+end subroutine
+
+! CHECK-LABEL: func.func @_QPtest_atomicadd_r4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
+! CHECK: llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f32
+
+attributes(global) subroutine test_atomicadd_r8()
+  real(8), device :: a, tmp1, tmp2
+  tmp1 = atomicadd(a, tmp2)
+end subroutine
+
+! CHECK-LABEL: func.func @_QPtest_atomicadd_r8() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
+! CHECK: llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f64
+
+attributes(global) subroutine test_atomicadd_i4()
+  integer(4), device :: a, tmp1, tmp2
+  tmp1 = atomicadd(a, tmp2)
+end subroutine
+
+! CHECK-LABEL: func.func @_QPtest_atomicadd_i4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>{{.*}}}
+! CHECK: llvm.atomicrmw add %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, i32
diff --git a/flang/test/Lower/CUDA/cuda-device-proc.cuf b/flang/test/Lower/CUDA/cuda-device-proc.cuf
index a1c3c33a76811d..b9081681569be9 100644
--- a/flang/test/Lower/CUDA/cuda-device-proc.cuf
+++ b/flang/test/Lower/CUDA/cuda-device-proc.cuf
@@ -123,9 +123,9 @@ end
 ! CHECK: %{{.*}} = nvvm.barrier.reduction #nvvm.reduction<or> %[[CONV]] -> i32
 ! CHECK: %{{.*}} = llvm.atomicrmw add  %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, i32
 ! CHECK: %{{.*}} = llvm.atomicrmw add  %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, i64
-! CHECK: %{{.*}} = llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f32
-! CHECK: %{{.*}} = llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f64
-! CHECK: %{{.*}} = llvm.atomicrmw fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf16>
+! CHECK: %{{.*}} = llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f32
+! CHECK: %{{.*}} = llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, f64
+! CHECK: %{{.*}} = llvm.atomicrmw ignore_denormal_mode fadd %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, vector<2xf16>
 
 ! CHECK: %{{.*}} = llvm.atomicrmw sub  %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, i32
 ! CHECK: %{{.*}} = llvm.atomicrmw sub  %{{.*}}, %{{.*}} seq_cst : !llvm.ptr, i64