125 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py2# RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=bdver2 -timeline -timeline-max-iterations=2 < %s | FileCheck %s3 4# VALU0/VALU15vpmulld %xmm0, %xmm1, %xmm26vpand %xmm0, %xmm1, %xmm27 8# VIMUL/STC9vcvttps2dq %xmm0, %xmm210vpclmulqdq $0, %xmm0, %xmm1, %xmm211 12# FPA/FPM13vaddps %xmm0, %xmm1, %xmm214vsqrtps %xmm0, %xmm215 16# FPA/FPM YMM17vaddps %ymm0, %ymm1, %ymm218vsqrtps %ymm0, %ymm219 20# CHECK: Iterations: 10021# CHECK-NEXT: Instructions: 80022# CHECK-NEXT: Total Cycles: 150323# CHECK-NEXT: Total uOps: 150024 25# CHECK: Dispatch Width: 426# CHECK-NEXT: uOps Per Cycle: 1.0027# CHECK-NEXT: IPC: 0.5328# CHECK-NEXT: Block RThroughput: 15.029 30# CHECK: Instruction Info:31# CHECK-NEXT: [1]: #uOps32# CHECK-NEXT: [2]: Latency33# CHECK-NEXT: [3]: RThroughput34# CHECK-NEXT: [4]: MayLoad35# CHECK-NEXT: [5]: MayStore36# 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, %xmm240# CHECK-NEXT: 1 2 0.50 vpand %xmm0, %xmm1, %xmm241# CHECK-NEXT: 1 4 1.00 vcvttps2dq %xmm0, %xmm242# CHECK-NEXT: 6 12 7.00 vpclmulqdq $0, %xmm0, %xmm1, %xmm243# CHECK-NEXT: 1 5 1.00 vaddps %xmm0, %xmm1, %xmm244# CHECK-NEXT: 1 9 4.50 vsqrtps %xmm0, %xmm245# CHECK-NEXT: 2 5 1.00 vaddps %ymm0, %ymm1, %ymm246# CHECK-NEXT: 2 9 9.00 vsqrtps %ymm0, %ymm247 48# CHECK: Resources:49# CHECK-NEXT: [0.0] - PdAGLU0150# CHECK-NEXT: [0.1] - PdAGLU0151# CHECK-NEXT: [1] - PdBranch52# CHECK-NEXT: [2] - PdCount53# CHECK-NEXT: [3] - PdDiv54# CHECK-NEXT: [4] - PdEX055# CHECK-NEXT: [5] - PdEX156# CHECK-NEXT: [6] - PdFPCVT57# CHECK-NEXT: [7.0] - PdFPFMA58# CHECK-NEXT: [7.1] - PdFPFMA59# CHECK-NEXT: [8.0] - PdFPMAL60# CHECK-NEXT: [8.1] - PdFPMAL61# CHECK-NEXT: [9] - PdFPMMA62# CHECK-NEXT: [10] - PdFPSTO63# CHECK-NEXT: [11] - PdFPU064# CHECK-NEXT: [12] - PdFPU165# CHECK-NEXT: [13] - PdFPU266# CHECK-NEXT: [14] - PdFPU367# CHECK-NEXT: [15] - PdFPXBR68# CHECK-NEXT: [16.0] - PdLoad69# CHECK-NEXT: [16.1] - PdLoad70# CHECK-NEXT: [17] - PdMul71# CHECK-NEXT: [18] - PdStore72 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 1.00 1.00 9.00 1.00 6.00 4.00 0.50 0.50 - - - - -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, %xmm280# CHECK-NEXT: - - - - - - - - - - 0.12 0.88 - - - - 0.50 0.50 - - - - - vpand %xmm0, %xmm1, %xmm281# CHECK-NEXT: - - - - - - - 1.00 - - - - - 1.00 1.00 - - - - - - - - vcvttps2dq %xmm0, %xmm282# CHECK-NEXT: - - - - - - - - - - - - 7.00 - 1.00 - - - - - - - - vpclmulqdq $0, %xmm0, %xmm1, %xmm283# CHECK-NEXT: - - - - - - - - 0.52 0.48 - - - - 1.00 - - - - - - - - vaddps %xmm0, %xmm1, %xmm284# CHECK-NEXT: - - - - - - - - 4.50 4.50 - - - - - 1.00 - - - - - - - vsqrtps %xmm0, %xmm285# CHECK-NEXT: - - - - - - - - 1.04 0.96 - - - - 1.00 - - - - - - - - vaddps %ymm0, %ymm1, %ymm286# CHECK-NEXT: - - - - - - - - 9.00 9.00 - - - - - 2.00 - - - - - - - vsqrtps %ymm0, %ymm287 88# CHECK: Timeline view:89# CHECK-NEXT: 0123456789 01290# CHECK-NEXT: Index 0123456789 012345678991 92# CHECK: [0,0] DeeeeeER . . . . . . vpmulld %xmm0, %xmm1, %xmm293# CHECK-NEXT: [0,1] DeeE---R . . . . . . vpand %xmm0, %xmm1, %xmm294# CHECK-NEXT: [0,2] D==eeeeER . . . . . . vcvttps2dq %xmm0, %xmm295# CHECK-NEXT: [0,3] .D==eeeeeeeeeeeeER . . . . vpclmulqdq $0, %xmm0, %xmm1, %xmm296# CHECK-NEXT: [0,4] . D===================eeeeeER . . vaddps %xmm0, %xmm1, %xmm297# CHECK-NEXT: [0,5] . DeeeeeeeeeE---------------R . . vsqrtps %xmm0, %xmm298# CHECK-NEXT: [0,6] . D===================eeeeeER. . vaddps %ymm0, %ymm1, %ymm299# CHECK-NEXT: [0,7] . DeeeeeeeeeE---------------R. . vsqrtps %ymm0, %ymm2100# CHECK-NEXT: [1,0] . D======eeeeeE------------R. . vpmulld %xmm0, %xmm1, %xmm2101# CHECK-NEXT: [1,1] . DeeE---------------------R. . vpand %xmm0, %xmm1, %xmm2102# CHECK-NEXT: [1,2] . DeeeeE--------------------R . vcvttps2dq %xmm0, %xmm2103# CHECK-NEXT: [1,3] . D=======eeeeeeeeeeeeE----R . vpclmulqdq $0, %xmm0, %xmm1, %xmm2104# CHECK-NEXT: [1,4] . .D==================eeeeeER. vaddps %xmm0, %xmm1, %xmm2105# CHECK-NEXT: [1,5] . .D=====eeeeeeeeeE---------R. vsqrtps %xmm0, %xmm2106# CHECK-NEXT: [1,6] . . D==================eeeeeER vaddps %ymm0, %ymm1, %ymm2107# CHECK-NEXT: [1,7] . . D=============eeeeeeeeeE-R vsqrtps %ymm0, %ymm2108 109# CHECK: Average Wait times (based on the timeline view):110# CHECK-NEXT: [0]: Executions111# CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue112# CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready113# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage114 115# CHECK: [0] [1] [2] [3]116# CHECK-NEXT: 0. 2 4.0 4.0 6.0 vpmulld %xmm0, %xmm1, %xmm2117# CHECK-NEXT: 1. 2 1.0 1.0 12.0 vpand %xmm0, %xmm1, %xmm2118# CHECK-NEXT: 2. 2 2.0 2.0 10.0 vcvttps2dq %xmm0, %xmm2119# CHECK-NEXT: 3. 2 5.5 5.5 2.0 vpclmulqdq $0, %xmm0, %xmm1, %xmm2120# CHECK-NEXT: 4. 2 19.5 19.5 0.0 vaddps %xmm0, %xmm1, %xmm2121# CHECK-NEXT: 5. 2 3.5 3.5 12.0 vsqrtps %xmm0, %xmm2122# CHECK-NEXT: 6. 2 19.5 19.5 0.0 vaddps %ymm0, %ymm1, %ymm2123# CHECK-NEXT: 7. 2 7.5 7.5 8.0 vsqrtps %ymm0, %ymm2124# CHECK-NEXT: 2 7.8 7.8 6.3 <total>125