[flang][cuda] Allow DO CONCURRENT in cuf kernel (#124190)
[llvm-project.git] / llvm / test / tools / llvm-mca / AArch64 / Cortex / A53-carry-over.s
bloba5715b965210982463a9f2787daac335dc7bb496
1 # NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
2 # RUN: llvm-mca -mtriple=aarch64 -mcpu=cortex-a53 --timeline --iterations=1 < %s | FileCheck %s
4 ldp w3, w5, [x10], #4 // 2uop + 1uop carry over
5 add w10, w11, w12
6 add w13, w14, w15
7 ldp w7, w8, [x11] // 2uop, no carry over
8 add w16, w17, w18
9 add w19, w20, w21
11 # CHECK: Iterations: 1
12 # CHECK-NEXT: Instructions: 6
13 # CHECK-NEXT: Total Cycles: 8
14 # CHECK-NEXT: Total uOps: 9
16 # CHECK: Dispatch Width: 2
17 # CHECK-NEXT: uOps Per Cycle: 1.13
18 # CHECK-NEXT: IPC: 0.75
19 # CHECK-NEXT: Block RThroughput: 4.5
21 # CHECK: Instruction Info:
22 # CHECK-NEXT: [1]: #uOps
23 # CHECK-NEXT: [2]: Latency
24 # CHECK-NEXT: [3]: RThroughput
25 # CHECK-NEXT: [4]: MayLoad
26 # CHECK-NEXT: [5]: MayStore
27 # CHECK-NEXT: [6]: HasSideEffects (U)
29 # CHECK: [1] [2] [3] [4] [5] [6] Instructions:
30 # CHECK-NEXT: 3 4 2.00 * ldp w3, w5, [x10], #4
31 # CHECK-NEXT: 1 3 0.50 add w10, w11, w12
32 # CHECK-NEXT: 1 3 0.50 add w13, w14, w15
33 # CHECK-NEXT: 2 4 2.00 * ldp w7, w8, [x11]
34 # CHECK-NEXT: 1 3 0.50 add w16, w17, w18
35 # CHECK-NEXT: 1 3 0.50 add w19, w20, w21
37 # CHECK: Resources:
38 # CHECK-NEXT: [0.0] - A53UnitALU
39 # CHECK-NEXT: [0.1] - A53UnitALU
40 # CHECK-NEXT: [1] - A53UnitB
41 # CHECK-NEXT: [2] - A53UnitDiv
42 # CHECK-NEXT: [3] - A53UnitFPALU
43 # CHECK-NEXT: [4] - A53UnitFPMDS
44 # CHECK-NEXT: [5] - A53UnitLdSt
45 # CHECK-NEXT: [6] - A53UnitMAC
47 # CHECK: Resource pressure per iteration:
48 # CHECK-NEXT: [0.0] [0.1] [1] [2] [3] [4] [5] [6]
49 # CHECK-NEXT: 2.00 2.00 - - - - 4.00 -
51 # CHECK: Resource pressure by instruction:
52 # CHECK-NEXT: [0.0] [0.1] [1] [2] [3] [4] [5] [6] Instructions:
53 # CHECK-NEXT: - - - - - - 2.00 - ldp w3, w5, [x10], #4
54 # CHECK-NEXT: - 1.00 - - - - - - add w10, w11, w12
55 # CHECK-NEXT: 1.00 - - - - - - - add w13, w14, w15
56 # CHECK-NEXT: - - - - - - 2.00 - ldp w7, w8, [x11]
57 # CHECK-NEXT: - 1.00 - - - - - - add w16, w17, w18
58 # CHECK-NEXT: 1.00 - - - - - - - add w19, w20, w21
60 # CHECK: Timeline view:
61 # CHECK-NEXT: Index 01234567
63 # CHECK: [0,0] DeeeE. . ldp w3, w5, [x10], #4
64 # CHECK-NEXT: [0,1] .DeeE. . add w10, w11, w12
65 # CHECK-NEXT: [0,2] . DeeE . add w13, w14, w15
66 # CHECK-NEXT: [0,3] . DeeeE ldp w7, w8, [x11]
67 # CHECK-NEXT: [0,4] . DeeE add w16, w17, w18
68 # CHECK-NEXT: [0,5] . DeeE add w19, w20, w21
70 # CHECK: Average Wait times (based on the timeline view):
71 # CHECK-NEXT: [0]: Executions
72 # CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue
73 # CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready
74 # CHECK-NEXT: [3]: Average time elapsed from WB until retire stage
76 # CHECK: [0] [1] [2] [3]
77 # CHECK-NEXT: 0. 1 0.0 0.0 0.0 ldp w3, w5, [x10], #4
78 # CHECK-NEXT: 1. 1 0.0 0.0 0.0 add w10, w11, w12
79 # CHECK-NEXT: 2. 1 0.0 0.0 0.0 add w13, w14, w15
80 # CHECK-NEXT: 3. 1 0.0 0.0 0.0 ldp w7, w8, [x11]
81 # CHECK-NEXT: 4. 1 0.0 0.0 0.0 add w16, w17, w18
82 # CHECK-NEXT: 5. 1 0.0 0.0 0.0 add w19, w20, w21
83 # CHECK-NEXT: 1 0.0 0.0 0.0 <total>