| // RUN: fir-opt --opt-bufferization %s | FileCheck %s | |
| // Verify that ElementalAssignBufferization does NOT fuse an elemental into the | |
| // assignment when a deallocation between the elemental and the assignment frees | |
| // memory that the elemental reads. Moving the elemental evaluation down to the | |
| // assignment would read freed memory. | |
| func.func @dealloc_conflict(%dst: !fir.box<!fir.array<?xf32>>, %n: index) { | |
| %c1 = arith.constant 1 : index | |
| %cst = arith.constant 1.000000e+00 : f32 | |
| %shape = fir.shape %n : (index) -> !fir.shape<1> | |
| %heap = fir.allocmem !fir.array<?xf32>, %n {uniq_name = ".src"} | |
| %src = fir.embox %heap(%shape) : (!fir.heap<!fir.array<?xf32>>, !fir.shape<1>) -> !fir.box<!fir.array<?xf32>> | |
| %elem = hlfir.elemental %shape unordered : (!fir.shape<1>) -> !hlfir.expr<?xf32> { | |
| ^bb0(%i: index): | |
| %d = hlfir.designate %src (%i) : (!fir.box<!fir.array<?xf32>>, index) -> !fir.ref<f32> | |
| %v = fir.load %d : !fir.ref<f32> | |
| %s = arith.addf %v, %cst fastmath<contract> : f32 | |
| hlfir.yield_element %s : f32 | |
| } | |
| // Deallocation of the elemental's read source between elemental and assign. | |
| fir.freemem %heap : !fir.heap<!fir.array<?xf32>> | |
| hlfir.assign %elem to %dst : !hlfir.expr<?xf32>, !fir.box<!fir.array<?xf32>> | |
| hlfir.destroy %elem : !hlfir.expr<?xf32> | |
| return | |
| } | |
| // CHECK-LABEL: func.func @dealloc_conflict | |
| // The elemental and the array-level assign must be preserved (no fusion). | |
| // CHECK: hlfir.elemental | |
| // CHECK: fir.freemem | |
| // CHECK: hlfir.assign %{{.*}} to %{{.*}} : !hlfir.expr<?xf32>, !fir.box<!fir.array<?xf32>> | |
| // Same conflict through cuf.free, which frees the memory of its operand. | |
| func.func @cuf_dealloc_conflict(%dst: !fir.box<!fir.array<?xf32>>, %n: index) { | |
| %cst = arith.constant 1.000000e+00 : f32 | |
| %shape = fir.shape %n : (index) -> !fir.shape<1> | |
| %ptr = cuf.alloc !fir.array<?xf32>, %n : index {bindc_name = ".src", data_attr = #cuf.cuda<managed>, uniq_name = ".src"} -> !fir.ref<!fir.array<?xf32>> | |
| %heap = fir.convert %ptr : (!fir.ref<!fir.array<?xf32>>) -> !fir.heap<!fir.array<?xf32>> | |
| %src = fir.embox %heap(%shape) : (!fir.heap<!fir.array<?xf32>>, !fir.shape<1>) -> !fir.box<!fir.array<?xf32>> | |
| %elem = hlfir.elemental %shape unordered : (!fir.shape<1>) -> !hlfir.expr<?xf32> { | |
| ^bb0(%i: index): | |
| %d = hlfir.designate %src (%i) : (!fir.box<!fir.array<?xf32>>, index) -> !fir.ref<f32> | |
| %v = fir.load %d : !fir.ref<f32> | |
| %s = arith.addf %v, %cst fastmath<contract> : f32 | |
| hlfir.yield_element %s : f32 | |
| } | |
| cuf.free %ptr : !fir.ref<!fir.array<?xf32>> {data_attr = #cuf.cuda<managed>} | |
| hlfir.assign %elem to %dst : !hlfir.expr<?xf32>, !fir.box<!fir.array<?xf32>> | |
| hlfir.destroy %elem : !hlfir.expr<?xf32> | |
| return | |
| } | |
| // CHECK-LABEL: func.func @cuf_dealloc_conflict | |
| // CHECK: hlfir.elemental | |
| // CHECK: cuf.free | |
| // CHECK: hlfir.assign %{{.*}} to %{{.*}} : !hlfir.expr<?xf32>, !fir.box<!fir.array<?xf32>> | |
| // A cuf.free of memory that the elemental does not read must not block the | |
| // fusion: the free effect is attached to the freed pointer, so the alias | |
| // analysis can tell the two apart. | |
| func.func @cuf_dealloc_no_conflict(%dst: !fir.box<!fir.array<?xf32>>, %src: !fir.box<!fir.array<?xf32>>, %n: index) { | |
| %cst = arith.constant 1.000000e+00 : f32 | |
| %shape = fir.shape %n : (index) -> !fir.shape<1> | |
| %ptr = cuf.alloc !fir.array<?xf32>, %n : index {bindc_name = ".unrelated", data_attr = #cuf.cuda<managed>, uniq_name = ".unrelated"} -> !fir.ref<!fir.array<?xf32>> | |
| %elem = hlfir.elemental %shape unordered : (!fir.shape<1>) -> !hlfir.expr<?xf32> { | |
| ^bb0(%i: index): | |
| %d = hlfir.designate %src (%i) : (!fir.box<!fir.array<?xf32>>, index) -> !fir.ref<f32> | |
| %v = fir.load %d : !fir.ref<f32> | |
| %s = arith.addf %v, %cst fastmath<contract> : f32 | |
| hlfir.yield_element %s : f32 | |
| } | |
| cuf.free %ptr : !fir.ref<!fir.array<?xf32>> {data_attr = #cuf.cuda<managed>} | |
| hlfir.assign %elem to %dst : !hlfir.expr<?xf32>, !fir.box<!fir.array<?xf32>> | |
| hlfir.destroy %elem : !hlfir.expr<?xf32> | |
| return | |
| } | |
| // CHECK-LABEL: func.func @cuf_dealloc_no_conflict | |
| // The elemental is fused into the assignment loop. | |
| // CHECK-NOT: hlfir.elemental | |
| // CHECK: fir.do_loop | |
| // CHECK: hlfir.assign %{{.*}} to %{{.*}} : f32, !fir.ref<f32> |