1# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py 2# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=bdver2 -timeline -timeline-max-iterations=2 < %s | FileCheck %s 3 4# VALU0/VALU1 5vpmulld %xmm0, %xmm1, %xmm2 6vpand %xmm0, %xmm1, %xmm2 7 8# VIMUL/STC 9vcvttps2dq %xmm0, %xmm2 10vpclmulqdq $0, %xmm0, %xmm1, %xmm2 11 12# FPA/FPM 13vaddps %xmm0, %xmm1, %xmm2 14vsqrtps %xmm0, %xmm2 15 16# FPA/FPM YMM 17vaddps %ymm0, %ymm1, %ymm2 18vsqrtps %ymm0, %ymm2 19 20# CHECK: Iterations: 100 21# CHECK-NEXT: Instructions: 800 22# CHECK-NEXT: Total Cycles: 1503 23# CHECK-NEXT: Total uOps: 1500 24 25# CHECK: Dispatch Width: 4 26# CHECK-NEXT: uOps Per Cycle: 1.00 27# CHECK-NEXT: IPC: 0.53 28# CHECK-NEXT: Block RThroughput: 15.0 29 30# CHECK: Instruction Info: 31# CHECK-NEXT: [1]: #uOps 32# CHECK-NEXT: [2]: Latency 33# CHECK-NEXT: [3]: RThroughput 34# CHECK-NEXT: [4]: MayLoad 35# CHECK-NEXT: [5]: MayStore 36# CHECK-NEXT: [6]: HasSideEffects (U) 37 38# CHECK: [1] [2] [3] [4] [5] [6] Instructions: 39# CHECK-NEXT: 1 5 2.00 vpmulld %xmm0, %xmm1, %xmm2 40# CHECK-NEXT: 1 2 1.00 vpand %xmm0, %xmm1, %xmm2 41# CHECK-NEXT: 1 4 1.00 vcvttps2dq %xmm0, %xmm2 42# CHECK-NEXT: 6 12 7.00 vpclmulqdq $0, %xmm0, %xmm1, %xmm2 43# CHECK-NEXT: 1 5 1.00 vaddps %xmm0, %xmm1, %xmm2 44# CHECK-NEXT: 1 9 4.50 vsqrtps %xmm0, %xmm2 45# CHECK-NEXT: 2 5 1.00 vaddps %ymm0, %ymm1, %ymm2 46# CHECK-NEXT: 2 9 9.00 vsqrtps %ymm0, %ymm2 47 48# CHECK: Resources: 49# CHECK-NEXT: [0.0] - PdAGLU01 50# CHECK-NEXT: [0.1] - PdAGLU01 51# CHECK-NEXT: [1] - PdBranch 52# CHECK-NEXT: [2] - PdCount 53# CHECK-NEXT: [3] - PdDiv 54# CHECK-NEXT: [4] - PdEX0 55# CHECK-NEXT: [5] - PdEX1 56# CHECK-NEXT: [6] - PdFPCVT 57# CHECK-NEXT: [7.0] - PdFPFMA 58# CHECK-NEXT: [7.1] - PdFPFMA 59# CHECK-NEXT: [8.0] - PdFPMAL 60# CHECK-NEXT: [8.1] - PdFPMAL 61# CHECK-NEXT: [9] - PdFPMMA 62# CHECK-NEXT: [10] - PdFPSTO 63# CHECK-NEXT: [11] - PdFPU0 64# CHECK-NEXT: [12] - PdFPU1 65# CHECK-NEXT: [13] - PdFPU2 66# CHECK-NEXT: [14] - PdFPU3 67# CHECK-NEXT: [15] - PdFPXBR 68# CHECK-NEXT: [16.0] - PdLoad 69# CHECK-NEXT: [16.1] - PdLoad 70# CHECK-NEXT: [17] - PdMul 71# CHECK-NEXT: [18] - PdStore 72 73# CHECK: Resource pressure per iteration: 74# CHECK-NEXT: [0.0] [0.1] [1] [2] [3] [4] [5] [6] [7.0] [7.1] [8.0] [8.1] [9] [10] [11] [12] [13] [14] [15] [16.0] [16.1] [17] [18] 75# CHECK-NEXT: - - - - - - - 1.00 15.06 14.94 1.12 1.88 9.00 1.00 6.44 4.56 - - - - - - - 76 77# CHECK: Resource pressure by instruction: 78# CHECK-NEXT: [0.0] [0.1] [1] [2] [3] [4] [5] [6] [7.0] [7.1] [8.0] [8.1] [9] [10] [11] [12] [13] [14] [15] [16.0] [16.1] [17] [18] Instructions: 79# CHECK-NEXT: - - - - - - - - - - 0.88 0.12 2.00 - 2.00 1.00 - - - - - - - vpmulld %xmm0, %xmm1, %xmm2 80# CHECK-NEXT: - - - - - - - - - - 0.24 1.76 - - 0.44 0.56 - - - - - - - vpand %xmm0, %xmm1, %xmm2 81# CHECK-NEXT: - - - - - - - 1.00 - - - - - 1.00 1.00 - - - - - - - - vcvttps2dq %xmm0, %xmm2 82# CHECK-NEXT: - - - - - - - - - - - - 7.00 - 1.00 - - - - - - - - vpclmulqdq $0, %xmm0, %xmm1, %xmm2 83# CHECK-NEXT: - - - - - - - - 0.52 0.48 - - - - 1.00 - - - - - - - - vaddps %xmm0, %xmm1, %xmm2 84# CHECK-NEXT: - - - - - - - - 4.50 4.50 - - - - - 1.00 - - - - - - - vsqrtps %xmm0, %xmm2 85# CHECK-NEXT: - - - - - - - - 1.04 0.96 - - - - 1.00 - - - - - - - - vaddps %ymm0, %ymm1, %ymm2 86# CHECK-NEXT: - - - - - - - - 9.00 9.00 - - - - - 2.00 - - - - - - - vsqrtps %ymm0, %ymm2 87 88# CHECK: Timeline view: 89# CHECK-NEXT: 0123456789 012 90# CHECK-NEXT: Index 0123456789 0123456789 91 92# CHECK: [0,0] DeeeeeER . . . . . . vpmulld %xmm0, %xmm1, %xmm2 93# CHECK-NEXT: [0,1] D=eeE--R . . . . . . vpand %xmm0, %xmm1, %xmm2 94# CHECK-NEXT: [0,2] D==eeeeER . . . . . . vcvttps2dq %xmm0, %xmm2 95# CHECK-NEXT: [0,3] .D==eeeeeeeeeeeeER . . . . vpclmulqdq $0, %xmm0, %xmm1, %xmm2 96# CHECK-NEXT: [0,4] . D===================eeeeeER . . vaddps %xmm0, %xmm1, %xmm2 97# CHECK-NEXT: [0,5] . DeeeeeeeeeE---------------R . . vsqrtps %xmm0, %xmm2 98# CHECK-NEXT: [0,6] . D===================eeeeeER. . vaddps %ymm0, %ymm1, %ymm2 99# CHECK-NEXT: [0,7] . DeeeeeeeeeE---------------R. . vsqrtps %ymm0, %ymm2 100# CHECK-NEXT: [1,0] . D======eeeeeE------------R. . vpmulld %xmm0, %xmm1, %xmm2 101# CHECK-NEXT: [1,1] . DeeE---------------------R. . vpand %xmm0, %xmm1, %xmm2 102# CHECK-NEXT: [1,2] . D=eeeeE-------------------R . vcvttps2dq %xmm0, %xmm2 103# CHECK-NEXT: [1,3] . D=======eeeeeeeeeeeeE----R . vpclmulqdq $0, %xmm0, %xmm1, %xmm2 104# CHECK-NEXT: [1,4] . .D==================eeeeeER. vaddps %xmm0, %xmm1, %xmm2 105# CHECK-NEXT: [1,5] . .D=====eeeeeeeeeE---------R. vsqrtps %xmm0, %xmm2 106# CHECK-NEXT: [1,6] . . D==================eeeeeER vaddps %ymm0, %ymm1, %ymm2 107# CHECK-NEXT: [1,7] . . D=============eeeeeeeeeE-R vsqrtps %ymm0, %ymm2 108 109# CHECK: Average Wait times (based on the timeline view): 110# CHECK-NEXT: [0]: Executions 111# CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue 112# CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready 113# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage 114 115# CHECK: [0] [1] [2] [3] 116# CHECK-NEXT: 0. 2 4.0 4.0 6.0 vpmulld %xmm0, %xmm1, %xmm2 117# CHECK-NEXT: 1. 2 1.5 1.5 11.5 vpand %xmm0, %xmm1, %xmm2 118# CHECK-NEXT: 2. 2 2.5 2.5 9.5 vcvttps2dq %xmm0, %xmm2 119# CHECK-NEXT: 3. 2 5.5 5.5 2.0 vpclmulqdq $0, %xmm0, %xmm1, %xmm2 120# CHECK-NEXT: 4. 2 19.5 19.5 0.0 vaddps %xmm0, %xmm1, %xmm2 121# CHECK-NEXT: 5. 2 3.5 3.5 12.0 vsqrtps %xmm0, %xmm2 122# CHECK-NEXT: 6. 2 19.5 19.5 0.0 vaddps %ymm0, %ymm1, %ymm2 123# CHECK-NEXT: 7. 2 7.5 7.5 8.0 vsqrtps %ymm0, %ymm2 124# CHECK-NEXT: 2 7.9 7.9 6.1 <total> 125