Skip to content

Commit 16c1fc7

Browse files
MacDueMacDue
committed
More stuff
1 parent 1373a8e commit 16c1fc7

File tree

1 file changed

+85
-8
lines changed

1 file changed

+85
-8
lines changed

‎mlir/test/Dialect/ArmSME/tile-allocation-copies.mlir

Lines changed: 85 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -43,20 +43,27 @@ func.func @cond_branch(%cond: i1, %tile: vector<[4]x[4]xf32>) {
4343
// Reduction of a real world example that shows why we must split conditional branches.
4444

4545
// CHECK-LABEL: @cond_branch_with_backedge(
46-
// CHECK-SAME: %{{[[:alnum:]]+}}: vector<[4]x[4]xf32>, %[[TILEB:[[:alnum:]]+]]: vector<[4]x[4]xf32>,
46+
// CHECK-SAME: %[[TILEA:[[:alnum:]]+]]: vector<[4]x[4]xf32>, %[[TILEB:[[:alnum:]]+]]: vector<[4]x[4]xf32>,
4747
// CHECK-SAME: %[[TILEC:[[:alnum:]]+]]: vector<[4]x[4]xf32>, %[[TILED:[[:alnum:]]+]]: vector<[4]x[4]xf32>,
48+
// CHECK: %[[BB1_COPY_0:.*]] = arm_sme.copy_tile %[[TILEA]] : vector<[4]x[4]xf32>
49+
// CHECK: cf.br ^bb1(%{{[[:alnum:]]+}}, %[[BB1_COPY_0]]
4850
// CHECK: ^bb1(%[[CURRENT_INDEX:.*]]: index, %[[ITER_TILE:.*]]: vector<[4]x[4]xf32>):
4951
// CHECK: %[[CONTINUE_LOOP:.*]] = arith.cmpi
50-
// CHECK: cf.cond_br %[[CONTINUE_LOOP]], ^[[BB2:[[:alnum:]]+]], ^[[BB3_COPIES:[[:alnum:]]+]]
52+
// CHECK: cf.cond_br %[[CONTINUE_LOOP]], ^[[BB2_COPIES:[[:alnum:]]+]], ^[[BB3_COPIES:[[:alnum:]]+]]
5153
// CHECK: ^[[BB3_COPIES]]:
52-
// CHECK-NEXT: arm_sme.copy_tile %[[ITER_TILE]] : vector<[4]x[4]xf32>
53-
// CHECK-NEXT: arm_sme.copy_tile %[[TILEB]] : vector<[4]x[4]xf32>
54-
// CHECK-NEXT: arm_sme.copy_tile %[[TILEC]] : vector<[4]x[4]xf32>
55-
// CHECK-NEXT: arm_sme.copy_tile %[[TILED]] : vector<[4]x[4]xf32>
56-
// CHECK-NEXT: cf.br ^[[BB3:[[:alnum:]]+]]
54+
// CHECK-NEXT: %[[BB3_COPY_0:.*]] = arm_sme.copy_tile %[[ITER_TILE]] : vector<[4]x[4]xf32>
55+
// CHECK-NEXT: %[[BB3_COPY_1:.*]] = arm_sme.copy_tile %[[TILEB]] : vector<[4]x[4]xf32>
56+
// CHECK-NEXT: %[[BB3_COPY_2:.*]] = arm_sme.copy_tile %[[TILEC]] : vector<[4]x[4]xf32>
57+
// CHECK-NEXT: %[[BB3_COPY_3:.*]] = arm_sme.copy_tile %[[TILED]] : vector<[4]x[4]xf32>
58+
// CHECK-NEXT: cf.br ^[[BB3:[[:alnum:]]+]](%[[BB3_COPY_0]], %[[BB3_COPY_1]], %[[BB3_COPY_2]], %[[BB3_COPY_3]]
59+
// CHECK: ^[[BB2_COPIES]]:
60+
// CHECK-NEXT: cf.br ^[[BB2:[[:alnum:]]+]]
61+
// CHECK: ^[[BB2]]:
62+
// CHECK-NEXT: %[[NEXT_TILE:.*]] = arm_sme.move_vector_to_tile_slice %{{.*}}, %[[ITER_TILE]]
63+
// CHECK: %[[BB1_COPY_1:.*]] = arm_sme.copy_tile %[[NEXT_TILE]] : vector<[4]x[4]xf32>
64+
// CHECK: cf.br ^bb1(%{{[[:alnum:]]+}}, %[[BB1_COPY_1]]
5765
// CHECK: ^[[BB3]](%{{.*}}: vector<[4]x[4]xf32>):
5866
// CHECK-NEXT: return
59-
6067
func.func @cond_branch_with_backedge(%tileA: vector<[4]x[4]xf32>, %tileB: vector<[4]x[4]xf32>, %tileC: vector<[4]x[4]xf32>, %tileD: vector<[4]x[4]xf32>, %slice: vector<[4]xf32>) {
6168
%c0 = arith.constant 0 : index
6269
%c1 = arith.constant 1 : index
@@ -80,3 +87,73 @@ func.func @cond_branch_with_backedge(%tileA: vector<[4]x[4]xf32>, %tileB: vector
8087
// Live here: %finalTileA, %finalTileB, %finalTileC, %finalTileD
8188
return
8289
}
90+
91+
// -----
92+
93+
// CHECK-LABEL: @tile_dominance
94+
// CHECK-NOT: arm_sme.copy_tile
95+
func.func @tile_dominance(%arg0: vector<[4]x[4]xf32>) {
96+
cf.br ^bb1
97+
^bb1: // 2 preds: ^bb0, ^bb4
98+
"test.some_use"(%arg0) : (vector<[4]x[4]xf32>) -> ()
99+
return
100+
^bb2: // no predecessors
101+
%0 = arm_sme.get_tile : vector<[4]x[4]xf32>
102+
cf.br ^bb3
103+
^bb3: // pred: ^bb2
104+
"test.some_use"(%0) : (vector<[4]x[4]xf32>) -> ()
105+
return
106+
^bb4: // no predecessors
107+
cf.br ^bb1
108+
^bb5: // no predecessors
109+
return
110+
}
111+
112+
// -----
113+
114+
// CHECK-LABEL: func.func @cond_branch_true_and_false_tile_args(
115+
// CHECK-SAME: %[[COND:.*]]: i1, %[[TILE:.*]]: vector<[4]x[4]xf32>
116+
// CHECK-NEXT: cf.cond_br %[[COND]], ^[[BB1_COPIES:[[:alnum:]]+]], ^[[BB2_COPIES:[[:alnum:]]+]]
117+
// CHECK: ^[[BB2_COPIES]]:
118+
// CHECK-NEXT: %[[COPY_0:.*]] = arm_sme.copy_tile %[[TILE]] : vector<[4]x[4]xf32>
119+
// CHECK-NEXT: cf.br ^[[BB2:[[:alnum:]]+]](%[[COPY_0]]
120+
// CHECK: ^[[BB1_COPIES]]:
121+
// CHECK-NEXT: %[[COPY_1:.*]] = arm_sme.copy_tile %[[TILE]] : vector<[4]x[4]xf32>
122+
// CHECK-NEXT: cf.br ^[[BB1:[[:alnum:]]+]](%[[COPY_1]]
123+
// CHECK: ^[[BB1]]{{.*}}:
124+
// CHECK-NEXT: return
125+
// CHECK: ^[[BB2]]{{.*}}:
126+
// CHECK-NEXT: return
127+
func.func @cond_branch_true_and_false_tile_args(%cond: i1, %tile: vector<[4]x[4]xf32>) {
128+
cf.cond_br %cond, ^bb1(%tile: vector<[4]x[4]xf32>), ^bb2(%tile: vector<[4]x[4]xf32>)
129+
^bb1(%blockArg0: vector<[4]x[4]xf32>):
130+
return
131+
^bb2(%blockArg1: vector<[4]x[4]xf32>):
132+
return
133+
}
134+
135+
// -----
136+
137+
// CHECK-LABEL: @multiple_predecessors
138+
// CHECK: ^bb1:
139+
// CHECK-NEXT: %[[TILE:.*]] = arm_sme.get_tile : vector<[4]x[4]xf32>
140+
// CHECK-NEXT: %[[COPY_0:.*]] = arm_sme.copy_tile %[[TILE]] : vector<[4]x[4]xf32>
141+
// CHECK-NEXT: cf.br ^bb3(%[[COPY_0]] : vector<[4]x[4]xf32>)
142+
// CHECK: ^bb2:
143+
// CHECK-NEXT: %[[ZERO:.*]] = arm_sme.zero : vector<[4]x[4]xf32>
144+
// CHECK-NEXT: %[[COPY_1:.*]] = arm_sme.copy_tile %[[ZERO]] : vector<[4]x[4]xf32>
145+
// CHECK-NEXT: cf.br ^bb3(%[[COPY_1]] : vector<[4]x[4]xf32>)
146+
// CHECK: ^bb3({{.*}}):
147+
// CHECK-NEXT: return
148+
func.func @multiple_predecessors(%cond: i1)
149+
{
150+
cf.cond_br %cond, ^bb1, ^bb2
151+
^bb1:
152+
%tile = arm_sme.get_tile : vector<[4]x[4]xf32>
153+
cf.br ^bb3(%tile : vector<[4]x[4]xf32>)
154+
^bb2:
155+
%zero = arm_sme.zero : vector<[4]x[4]xf32>
156+
cf.br ^bb3(%zero : vector<[4]x[4]xf32>)
157+
^bb3(%blockArg: vector<[4]x[4]xf32>): // pred: ^bb1, ^bb2
158+
return
159+
}

0 commit comments

Comments
 (0)