Skip to content

Commit 0106585

Browse files
ranapratap55ranapratap55
committed
Revert "[flang][OpenMP] Allocate reduction init temps on the stack for GPUs (llvm#146667)"
This reverts commit 8c9e0c6.
1 parent cb286da commit 0106585

File tree

2 files changed

+76
-88
lines changed

2 files changed

+76
-88
lines changed

‎flang/test/Lower/OpenMP/parallel-reduction-array.f90‎

Lines changed: 76 additions & 83 deletions
Original file line numberDiff line numberDiff line change
@@ -14,88 +14,81 @@ program reduce
1414
print *,i
1515
end program
1616

17-
! CPU-LABEL: omp.declare_reduction @add_reduction_byref_box_3xi32 : !fir.ref<!fir.box<!fir.array<3xi32>>> alloc {
18-
! CPU: %[[VAL_8:.*]] = fir.alloca !fir.box<!fir.array<3xi32>>
19-
! CPU: omp.yield(%[[VAL_8]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
20-
! CPU-LABEL: } init {
21-
! CPU: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>, %[[ALLOC:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
22-
! CPU: %[[VAL_2:.*]] = arith.constant 0 : i32
23-
! CPU: %[[VAL_3:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
24-
! CPU: %[[VAL_4:.*]] = arith.constant 3 : index
25-
! CPU: %[[VAL_5:.*]] = fir.shape %[[VAL_4]] : (index) -> !fir.shape<1>
26-
! CPU: %[[VAL_1:.*]] = fir.allocmem !fir.array<3xi32> {bindc_name = ".tmp", uniq_name = ""}
27-
! CPU: %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_1]](%[[VAL_5]]) {uniq_name = ".tmp"} : (!fir.heap<!fir.array<3xi32>>,
28-
! CPU: %[[TRUE:.*]] = arith.constant true
17+
! CHECK-LABEL: omp.declare_reduction @add_reduction_byref_box_3xi32 : !fir.ref<!fir.box<!fir.array<3xi32>>> alloc {
18+
! CHECK: %[[VAL_8:.*]] = fir.alloca !fir.box<!fir.array<3xi32>>
19+
! CHECK: omp.yield(%[[VAL_8]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
20+
! CHECK-LABEL: } init {
21+
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>, %[[ALLOC:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
22+
! CHECK: %[[VAL_2:.*]] = arith.constant 0 : i32
23+
! CHECK: %[[VAL_3:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
24+
! CHECK: %[[VAL_4:.*]] = arith.constant 3 : index
25+
! CHECK: %[[VAL_5:.*]] = fir.shape %[[VAL_4]] : (index) -> !fir.shape<1>
26+
! CHECK: %[[VAL_1:.*]] = fir.allocmem !fir.array<3xi32> {bindc_name = ".tmp", uniq_name = ""}
27+
! CHECK: %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_1]](%[[VAL_5]]) {uniq_name = ".tmp"} : (!fir.heap<!fir.array<3xi32>>,
28+
! CHECK: %[[TRUE:.*]] = arith.constant true
2929
!fir.shape<1>) -> (!fir.heap<!fir.array<3xi32>>, !fir.heap<!fir.array<3xi32>>)
30-
! CPU: %[[C0:.*]] = arith.constant 0 : index
31-
! CPU: %[[DIMS:.*]]:3 = fir.box_dims %[[VAL_3]], %[[C0]] : (!fir.box<!fir.array<3xi32>>, index) -> (index, index, index)
32-
! CPU: %[[SHIFT:.*]] = fir.shape_shift %[[DIMS]]#0, %[[DIMS]]#1 : (index, index) -> !fir.shapeshift<1>
33-
! CPU: %[[VAL_7:.*]] = fir.embox %[[VAL_6]]#0(%[[SHIFT]]) : (!fir.heap<!fir.array<3xi32>>, !fir.shapeshift<1>) -> !fir.box<!fir.array<3xi32>>
34-
! CPU: hlfir.assign %[[VAL_2]] to %[[VAL_7]] : i32, !fir.box<!fir.array<3xi32>>
35-
! CPU: fir.store %[[VAL_7]] to %[[ALLOC]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
36-
! CPU: omp.yield(%[[ALLOC]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
37-
! CPU: } combiner {
38-
! CPU: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>, %[[VAL_1:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
39-
! CPU: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
40-
! CPU: %[[VAL_3:.*]] = fir.load %[[VAL_1]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
41-
! CPU: %[[C1:.*]] = arith.constant 1 : index
42-
! CPU: %[[C3:.*]] = arith.constant 3 : index
43-
! CPU: %[[SHAPE_SHIFT:.*]] = fir.shape_shift %[[C1]], %[[C3]] : (index, index) -> !fir.shapeshift<1>
44-
! CPU: %[[C1_0:.*]] = arith.constant 1 : index
45-
! CPU: fir.do_loop %[[VAL_8:.*]] = %[[C1_0]] to %[[C3]] step %[[C1_0]] unordered {
46-
! CPU: %[[VAL_9:.*]] = fir.array_coor %[[VAL_2]](%[[SHAPE_SHIFT]]) %[[VAL_8]] : (!fir.box<!fir.array<3xi32>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
47-
! CPU: %[[VAL_10:.*]] = fir.array_coor %[[VAL_3]](%[[SHAPE_SHIFT]]) %[[VAL_8]] : (!fir.box<!fir.array<3xi32>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
48-
! CPU: %[[VAL_11:.*]] = fir.load %[[VAL_9]] : !fir.ref<i32>
49-
! CPU: %[[VAL_12:.*]] = fir.load %[[VAL_10]] : !fir.ref<i32>
50-
! CPU: %[[VAL_13:.*]] = arith.addi %[[VAL_11]], %[[VAL_12]] : i32
51-
! CPU: fir.store %[[VAL_13]] to %[[VAL_9]] : !fir.ref<i32>
52-
! CPU: }
53-
! CPU: omp.yield(%[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
54-
! CPU: } cleanup {
55-
! CPU: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
56-
! CPU: %[[VAL_1:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
57-
! CPU: %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.array<3xi32>>) -> !fir.ref<!fir.array<3xi32>>
58-
! CPU: %[[VAL_3:.*]] = fir.convert %[[VAL_2]] : (!fir.ref<!fir.array<3xi32>>) -> i64
59-
! CPU: %[[VAL_4:.*]] = arith.constant 0 : i64
60-
! CPU: %[[VAL_5:.*]] = arith.cmpi ne, %[[VAL_3]], %[[VAL_4]] : i64
61-
! CPU: fir.if %[[VAL_5]] {
62-
! CPU: %[[VAL_6:.*]] = fir.convert %[[VAL_2]] : (!fir.ref<!fir.array<3xi32>>) -> !fir.heap<!fir.array<3xi32>>
63-
! CPU: fir.freemem %[[VAL_6]] : !fir.heap<!fir.array<3xi32>>
64-
! CPU: }
65-
! CPU: omp.yield
66-
! CPU: }
30+
! CHECK: %[[C0:.*]] = arith.constant 0 : index
31+
! CHECK: %[[DIMS:.*]]:3 = fir.box_dims %[[VAL_3]], %[[C0]] : (!fir.box<!fir.array<3xi32>>, index) -> (index, index, index)
32+
! CHECK: %[[SHIFT:.*]] = fir.shape_shift %[[DIMS]]#0, %[[DIMS]]#1 : (index, index) -> !fir.shapeshift<1>
33+
! CHECK: %[[VAL_7:.*]] = fir.embox %[[VAL_6]]#0(%[[SHIFT]]) : (!fir.heap<!fir.array<3xi32>>, !fir.shapeshift<1>) -> !fir.box<!fir.array<3xi32>>
34+
! CHECK: hlfir.assign %[[VAL_2]] to %[[VAL_7]] : i32, !fir.box<!fir.array<3xi32>>
35+
! CHECK: fir.store %[[VAL_7]] to %[[ALLOC]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
36+
! CHECK: omp.yield(%[[ALLOC]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
37+
! CHECK: } combiner {
38+
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>, %[[VAL_1:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
39+
! CHECK: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
40+
! CHECK: %[[VAL_3:.*]] = fir.load %[[VAL_1]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
41+
! CHECK: %[[C1:.*]] = arith.constant 1 : index
42+
! CHECK: %[[C3:.*]] = arith.constant 3 : index
43+
! CHECK: %[[SHAPE_SHIFT:.*]] = fir.shape_shift %[[C1]], %[[C3]] : (index, index) -> !fir.shapeshift<1>
44+
! CHECK: %[[C1_0:.*]] = arith.constant 1 : index
45+
! CHECK: fir.do_loop %[[VAL_8:.*]] = %[[C1_0]] to %[[C3]] step %[[C1_0]] unordered {
46+
! CHECK: %[[VAL_9:.*]] = fir.array_coor %[[VAL_2]](%[[SHAPE_SHIFT]]) %[[VAL_8]] : (!fir.box<!fir.array<3xi32>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
47+
! CHECK: %[[VAL_10:.*]] = fir.array_coor %[[VAL_3]](%[[SHAPE_SHIFT]]) %[[VAL_8]] : (!fir.box<!fir.array<3xi32>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
48+
! CHECK: %[[VAL_11:.*]] = fir.load %[[VAL_9]] : !fir.ref<i32>
49+
! CHECK: %[[VAL_12:.*]] = fir.load %[[VAL_10]] : !fir.ref<i32>
50+
! CHECK: %[[VAL_13:.*]] = arith.addi %[[VAL_11]], %[[VAL_12]] : i32
51+
! CHECK: fir.store %[[VAL_13]] to %[[VAL_9]] : !fir.ref<i32>
52+
! CHECK: }
53+
! CHECK: omp.yield(%[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>)
54+
! CHECK: } cleanup {
55+
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<3xi32>>>):
56+
! CHECK: %[[VAL_1:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
57+
! CHECK: %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.array<3xi32>>) -> !fir.ref<!fir.array<3xi32>>
58+
! CHECK: %[[VAL_3:.*]] = fir.convert %[[VAL_2]] : (!fir.ref<!fir.array<3xi32>>) -> i64
59+
! CHECK: %[[VAL_4:.*]] = arith.constant 0 : i64
60+
! CHECK: %[[VAL_5:.*]] = arith.cmpi ne, %[[VAL_3]], %[[VAL_4]] : i64
61+
! CHECK: fir.if %[[VAL_5]] {
62+
! CHECK: %[[VAL_6:.*]] = fir.convert %[[VAL_2]] : (!fir.ref<!fir.array<3xi32>>) -> !fir.heap<!fir.array<3xi32>>
63+
! CHECK: fir.freemem %[[VAL_6]] : !fir.heap<!fir.array<3xi32>>
64+
! CHECK: }
65+
! CHECK: omp.yield
66+
! CHECK: }
6767

68-
! CPU-LABEL: func.func @_QQmain()
69-
! CPU: %[[VAL_0:.*]] = fir.address_of(@_QFEi) : !fir.ref<!fir.array<3xi32>>
70-
! CPU: %[[VAL_1:.*]] = arith.constant 3 : index
71-
! CPU: %[[VAL_2:.*]] = fir.shape %[[VAL_1]] : (index) -> !fir.shape<1>
72-
! CPU: %[[VAL_3:.*]]:2 = hlfir.declare %[[VAL_0]](%[[VAL_2]]) {uniq_name = "_QFEi"} : (!fir.ref<!fir.array<3xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<3xi32>>, !fir.ref<!fir.array<3xi32>>)
73-
! CPU: %[[VAL_4:.*]] = fir.embox %[[VAL_3]]#0(%[[VAL_2]]) : (!fir.ref<!fir.array<3xi32>>, !fir.shape<1>) -> !fir.box<!fir.array<3xi32>>
74-
! CPU: %[[VAL_5:.*]] = fir.alloca !fir.box<!fir.array<3xi32>>
75-
! CPU: fir.store %[[VAL_4]] to %[[VAL_5]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
76-
! CPU: omp.parallel reduction(byref @add_reduction_byref_box_3xi32 %[[VAL_5]] -> %[[VAL_6:.*]] : !fir.ref<!fir.box<!fir.array<3xi32>>>) {
77-
! CPU: %[[VAL_7:.*]]:2 = hlfir.declare %[[VAL_6]] {uniq_name = "_QFEi"} : (!fir.ref<!fir.box<!fir.array<3xi32>>>) -> (!fir.ref<!fir.box<!fir.array<3xi32>>>, !fir.ref<!fir.box<!fir.array<3xi32>>>)
78-
! CPU: %[[VAL_8:.*]] = arith.constant 1 : i32
79-
! CPU: %[[VAL_9:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
80-
! CPU: %[[VAL_10:.*]] = arith.constant 1 : index
81-
! CPU: %[[VAL_11:.*]] = hlfir.designate %[[VAL_9]] (%[[VAL_10]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
82-
! CPU: hlfir.assign %[[VAL_8]] to %[[VAL_11]] : i32, !fir.ref<i32>
83-
! CPU: %[[VAL_12:.*]] = arith.constant 2 : i32
84-
! CPU: %[[VAL_13:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
85-
! CPU: %[[VAL_14:.*]] = arith.constant 2 : index
86-
! CPU: %[[VAL_15:.*]] = hlfir.designate %[[VAL_13]] (%[[VAL_14]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
87-
! CPU: hlfir.assign %[[VAL_12]] to %[[VAL_15]] : i32, !fir.ref<i32>
88-
! CPU: %[[VAL_16:.*]] = arith.constant 3 : i32
89-
! CPU: %[[VAL_17:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
90-
! CPU: %[[VAL_18:.*]] = arith.constant 3 : index
91-
! CPU: %[[VAL_19:.*]] = hlfir.designate %[[VAL_17]] (%[[VAL_18]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
92-
! CPU: hlfir.assign %[[VAL_16]] to %[[VAL_19]] : i32, !fir.ref<i32>
93-
! CPU: omp.terminator
94-
! CPU: }
95-
96-
! GPU: omp.declare_reduction {{.*}} alloc {
97-
! GPU: } init {
98-
! GPU-NOT: fir.allocmem {{.*}} {bindc_name = ".tmp", {{.*}}}
99-
! GPU: fir.alloca {{.*}} {bindc_name = ".tmp"}
100-
! GPU: } combiner {
101-
! GPU: }
68+
! CHECK-LABEL: func.func @_QQmain()
69+
! CHECK: %[[VAL_0:.*]] = fir.address_of(@_QFEi) : !fir.ref<!fir.array<3xi32>>
70+
! CHECK: %[[VAL_1:.*]] = arith.constant 3 : index
71+
! CHECK: %[[VAL_2:.*]] = fir.shape %[[VAL_1]] : (index) -> !fir.shape<1>
72+
! CHECK: %[[VAL_3:.*]]:2 = hlfir.declare %[[VAL_0]](%[[VAL_2]]) {uniq_name = "_QFEi"} : (!fir.ref<!fir.array<3xi32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<3xi32>>, !fir.ref<!fir.array<3xi32>>)
73+
! CHECK: %[[VAL_4:.*]] = fir.embox %[[VAL_3]]#0(%[[VAL_2]]) : (!fir.ref<!fir.array<3xi32>>, !fir.shape<1>) -> !fir.box<!fir.array<3xi32>>
74+
! CHECK: %[[VAL_5:.*]] = fir.alloca !fir.box<!fir.array<3xi32>>
75+
! CHECK: fir.store %[[VAL_4]] to %[[VAL_5]] : !fir.ref<!fir.box<!fir.array<3xi32>>>
76+
! CHECK: omp.parallel reduction(byref @add_reduction_byref_box_3xi32 %[[VAL_5]] -> %[[VAL_6:.*]] : !fir.ref<!fir.box<!fir.array<3xi32>>>) {
77+
! CHECK: %[[VAL_7:.*]]:2 = hlfir.declare %[[VAL_6]] {uniq_name = "_QFEi"} : (!fir.ref<!fir.box<!fir.array<3xi32>>>) -> (!fir.ref<!fir.box<!fir.array<3xi32>>>, !fir.ref<!fir.box<!fir.array<3xi32>>>)
78+
! CHECK: %[[VAL_8:.*]] = arith.constant 1 : i32
79+
! CHECK: %[[VAL_9:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
80+
! CHECK: %[[VAL_10:.*]] = arith.constant 1 : index
81+
! CHECK: %[[VAL_11:.*]] = hlfir.designate %[[VAL_9]] (%[[VAL_10]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
82+
! CHECK: hlfir.assign %[[VAL_8]] to %[[VAL_11]] : i32, !fir.ref<i32>
83+
! CHECK: %[[VAL_12:.*]] = arith.constant 2 : i32
84+
! CHECK: %[[VAL_13:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
85+
! CHECK: %[[VAL_14:.*]] = arith.constant 2 : index
86+
! CHECK: %[[VAL_15:.*]] = hlfir.designate %[[VAL_13]] (%[[VAL_14]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
87+
! CHECK: hlfir.assign %[[VAL_12]] to %[[VAL_15]] : i32, !fir.ref<i32>
88+
! CHECK: %[[VAL_16:.*]] = arith.constant 3 : i32
89+
! CHECK: %[[VAL_17:.*]] = fir.load %[[VAL_7]]#0 : !fir.ref<!fir.box<!fir.array<3xi32>>>
90+
! CHECK: %[[VAL_18:.*]] = arith.constant 3 : index
91+
! CHECK: %[[VAL_19:.*]] = hlfir.designate %[[VAL_17]] (%[[VAL_18]]) : (!fir.box<!fir.array<3xi32>>, index) -> !fir.ref<i32>
92+
! CHECK: hlfir.assign %[[VAL_16]] to %[[VAL_19]] : i32, !fir.ref<i32>
93+
! CHECK: omp.terminator
94+
! CHECK: }

‎mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp‎

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1164,11 +1164,6 @@ initReductionVars(OP op, ArrayRef<BlockArgument> reductionArgs,
11641164
mapInitializationArgs(op, moduleTranslation, reductionDecls,
11651165
reductionVariableMap, i);
11661166

1167-
// TODO In some cases (specially on the GPU), the init regions may
1168-
// contains stack alloctaions. If the region is inlined in a loop, this is
1169-
// problematic. Instead of just inlining the region, handle allocations by
1170-
// hoisting fixed length allocations to the function entry and using
1171-
// stacksave and restore for variable length ones.
11721167
if (failed(inlineConvertOmpRegions(reductionDecls[i].getInitializerRegion(),
11731168
"omp.reduction.neutral", builder,
11741169
moduleTranslation, &phis)))

0 commit comments

Comments
 (0)