115 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 -iterations=100 -resource-pressure=false -timeline -timeline-max-iterations=2 < %s | FileCheck %s3 4# movss/movsd explicitly zeroes out the high bits of xmm,5# so addps can start immediately, without waiting for sqrtss to finish.6# AMD SOG for the AMD family 15h processors, 5.5 Partial-Register Writes7 8# LLVM-MCA-BEGIN9sqrtss %xmm0, %xmm010movss (%eax), %xmm011addps %xmm0, %xmm012# LLVM-MCA-END13 14# LLVM-MCA-BEGIN15sqrtsd %xmm0, %xmm016movsd (%eax), %xmm017addps %xmm0, %xmm018# LLVM-MCA-END19 20# CHECK: [0] Code Region21 22# CHECK: Iterations: 10023# CHECK-NEXT: Instructions: 30024# CHECK-NEXT: Total Cycles: 65525# CHECK-NEXT: Total uOps: 30026 27# CHECK: Dispatch Width: 428# CHECK-NEXT: uOps Per Cycle: 0.4629# CHECK-NEXT: IPC: 0.4630# CHECK-NEXT: Block RThroughput: 6.531 32# CHECK: Instruction Info:33# CHECK-NEXT: [1]: #uOps34# CHECK-NEXT: [2]: Latency35# CHECK-NEXT: [3]: RThroughput36# CHECK-NEXT: [4]: MayLoad37# CHECK-NEXT: [5]: MayStore38# CHECK-NEXT: [6]: HasSideEffects (U)39 40# CHECK: [1] [2] [3] [4] [5] [6] Instructions:41# CHECK-NEXT: 1 9 4.50 sqrtss %xmm0, %xmm042# CHECK-NEXT: 1 5 1.50 * movss (%eax), %xmm043# CHECK-NEXT: 1 5 1.00 addps %xmm0, %xmm044 45# CHECK: Timeline view:46# CHECK-NEXT: 012345678947# CHECK-NEXT: Index 0123456789 01248 49# CHECK: [0,0] D=eeeeeeeeeER . . . sqrtss %xmm0, %xmm050# CHECK-NEXT: [0,1] DeeeeeE-----R . . . movss (%eax), %xmm051# CHECK-NEXT: [0,2] D======eeeeeER . . . addps %xmm0, %xmm052# CHECK-NEXT: [1,0] D===========eeeeeeeeeER sqrtss %xmm0, %xmm053# CHECK-NEXT: [1,1] .D==eeeeeE------------R movss (%eax), %xmm054# CHECK-NEXT: [1,2] .D=========eeeeeE-----R addps %xmm0, %xmm055 56# CHECK: Average Wait times (based on the timeline view):57# CHECK-NEXT: [0]: Executions58# CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue59# CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready60# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage61 62# CHECK: [0] [1] [2] [3]63# CHECK-NEXT: 0. 2 7.0 1.0 0.0 sqrtss %xmm0, %xmm064# CHECK-NEXT: 1. 2 2.0 2.0 8.5 movss (%eax), %xmm065# CHECK-NEXT: 2. 2 8.5 1.5 2.5 addps %xmm0, %xmm066# CHECK-NEXT: 2 5.8 1.5 3.7 <total>67 68# CHECK: [1] Code Region69 70# CHECK: Iterations: 10071# CHECK-NEXT: Instructions: 30072# CHECK-NEXT: Total Cycles: 65573# CHECK-NEXT: Total uOps: 30074 75# CHECK: Dispatch Width: 476# CHECK-NEXT: uOps Per Cycle: 0.4677# CHECK-NEXT: IPC: 0.4678# CHECK-NEXT: Block RThroughput: 6.579 80# CHECK: Instruction Info:81# CHECK-NEXT: [1]: #uOps82# CHECK-NEXT: [2]: Latency83# CHECK-NEXT: [3]: RThroughput84# CHECK-NEXT: [4]: MayLoad85# CHECK-NEXT: [5]: MayStore86# CHECK-NEXT: [6]: HasSideEffects (U)87 88# CHECK: [1] [2] [3] [4] [5] [6] Instructions:89# CHECK-NEXT: 1 9 4.50 sqrtsd %xmm0, %xmm090# CHECK-NEXT: 1 5 1.50 * movsd (%eax), %xmm091# CHECK-NEXT: 1 5 1.00 addps %xmm0, %xmm092 93# CHECK: Timeline view:94# CHECK-NEXT: 012345678995# CHECK-NEXT: Index 0123456789 01296 97# CHECK: [0,0] D=eeeeeeeeeER . . . sqrtsd %xmm0, %xmm098# CHECK-NEXT: [0,1] DeeeeeE-----R . . . movsd (%eax), %xmm099# CHECK-NEXT: [0,2] D======eeeeeER . . . addps %xmm0, %xmm0100# CHECK-NEXT: [1,0] D===========eeeeeeeeeER sqrtsd %xmm0, %xmm0101# CHECK-NEXT: [1,1] .D==eeeeeE------------R movsd (%eax), %xmm0102# CHECK-NEXT: [1,2] .D=========eeeeeE-----R addps %xmm0, %xmm0103 104# CHECK: Average Wait times (based on the timeline view):105# CHECK-NEXT: [0]: Executions106# CHECK-NEXT: [1]: Average time spent waiting in a scheduler's queue107# CHECK-NEXT: [2]: Average time spent waiting in a scheduler's queue while ready108# CHECK-NEXT: [3]: Average time elapsed from WB until retire stage109 110# CHECK: [0] [1] [2] [3]111# CHECK-NEXT: 0. 2 7.0 1.0 0.0 sqrtsd %xmm0, %xmm0112# CHECK-NEXT: 1. 2 2.0 2.0 8.5 movsd (%eax), %xmm0113# CHECK-NEXT: 2. 2 8.5 1.5 2.5 addps %xmm0, %xmm0114# CHECK-NEXT: 2 5.8 1.5 3.7 <total>115