[flang][cuda] Handle floats in atomiccas

[clementval]

No description provided.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/128970.diff

2 Files Affected:

• (modified) flang/lib/Optimizer/Builder/IntrinsicCall.cpp (+14)
• (modified) flang/test/Lower/CUDA/cuda-device-proc.cuf (+23)

diff --git a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
index 537c817e32ad8..9a10ce949290a 100644
--- a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
@@ -2734,6 +2734,20 @@ mlir::Value IntrinsicLibrary::genAtomicCas(mlir::Type resultType,
 
   mlir::Value arg1 = args[1];
   mlir::Value arg2 = args[2];
+
+  auto bitCastFloat = [&](mlir::Value arg) -> mlir::Value {
+    if (mlir::isa<mlir::Float32Type>(arg.getType()))
+      return builder.create<mlir::LLVM::BitcastOp>(loc, builder.getI32Type(),
+                                                   arg);
+    if (mlir::isa<mlir::Float64Type>(arg.getType()))
+      return builder.create<mlir::LLVM::BitcastOp>(loc, builder.getI64Type(),
+                                                   arg);
+    return arg;
+  };
+
+  arg1 = bitCastFloat(arg1);
+  arg2 = bitCastFloat(arg2);
+
   if (arg1.getType() != arg2.getType()) {
     // arg1 and arg2 need to have the same type in AtomicCmpXchgOp.
     arg2 = builder.createConvert(loc, arg1.getType(), arg2);
diff --git a/flang/test/Lower/CUDA/cuda-device-proc.cuf b/flang/test/Lower/CUDA/cuda-device-proc.cuf
index f2b4eb57ad555..c651d34c55093 100644
--- a/flang/test/Lower/CUDA/cuda-device-proc.cuf
+++ b/flang/test/Lower/CUDA/cuda-device-proc.cuf
@@ -175,3 +175,26 @@ end subroutine
 ! CHECK: %[[VAL:.*]] = fir.convert %c14{{.*}} : (i32) -> i64
 ! CHECK: %[[ADDR:.*]] = builtin.unrealized_conversion_cast %{{.*}}#1 : !fir.ref<i64> to !llvm.ptr
 ! CHECK: llvm.cmpxchg %{{.*}}, %{{.*}}, %[[VAL]] acq_rel monotonic : !llvm.ptr, i64
+
+attributes(device) subroutine testAtomic3()
+  real :: a, i, istat
+  istat = atomiccas(a, i, 14.0)
+end subroutine
+
+! CHECK-LABEL: func.func @_QPtestatomic3()
+! CHECK: %[[BCAST1:.*]] = llvm.bitcast %{{.*}} : f32 to i32
+! CHECK: %[[BCAST2:.*]] = llvm.bitcast %{{.*}} : f32 to i32
+! CHECK: %[[CAST:.*]] = builtin.unrealized_conversion_cast %{{.*}}#1 : !fir.ref<f32> to !llvm.ptr
+! CHECK: llvm.cmpxchg %[[CAST]], %[[BCAST1]], %[[BCAST2]] acq_rel monotonic : !llvm.ptr, i32
+
+attributes(device) subroutine testAtomic4()
+  real(8) :: a, i, istat
+  istat = atomiccas(a, i, 14.0d0)
+end subroutine
+
+! CHECK-LABEL: func.func @_QPtestatomic4()
+! CHECK: %[[BCAST1:.*]] = llvm.bitcast %{{.*}} : f64 to i64
+! CHECK: %[[BCAST2:.*]] = llvm.bitcast %{{.*}} : f64 to i64
+! CHECK: %[[CAST:.*]] = builtin.unrealized_conversion_cast %{{.*}}#1 : !fir.ref<f64> to !llvm.ptr
+! CHECK: %[[ATOMIC:.*]] = llvm.cmpxchg %[[CAST]], %[[BCAST1]], %[[BCAST2]] acq_rel monotonic : !llvm.ptr, i64
+! CHECK: %[[RES:.*]] = llvm.extractvalue %[[ATOMIC]][1] : !llvm.struct<(i64, i1)>