llvm/test/tools/llvm-mca/X86/SapphireRapids/resources-avx512bitalg.s

   1 # NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
   2 # RUN: llvm-mca -mtriple=x86_64-unknown-unknown -mcpu=sapphirerapids -instruction-tables < %s | FileCheck %s
   3
   4 vpopcntb        %zmm1, %zmm0
   5 vpopcntb        (%rdi), %zmm0
   6 vpopcntb        %zmm1, %zmm0 {%k1}
   7 vpopcntb        (%rdi), %zmm0 {%k1}
   8 vpopcntb        %zmm1, %zmm0 {%k1} {z}
   9 vpopcntb        (%rdi), %zmm0 {%k1} {z}
  10
  11 vpopcntw        %zmm1, %zmm0
  12 vpopcntw        (%rdi), %zmm0
  13 vpopcntw        %zmm1, %zmm0 {%k1}
  14 vpopcntw        (%rdi), %zmm0 {%k1}
  15 vpopcntw        %zmm1, %zmm0 {%k1} {z}
  16 vpopcntw        (%rdi), %zmm0 {%k1} {z}
  17
  18 vpshufbitqmb    %zmm16, %zmm17, %k2
  19 vpshufbitqmb    (%rdi), %zmm17, %k2
  20 vpshufbitqmb    %zmm16, %zmm17, %k2 {%k1}
  21 vpshufbitqmb    (%rdi), %zmm17, %k2 {%k1}
  22
  23 # CHECK:      Instruction Info:
  24 # CHECK-NEXT: [1]: #uOps
  25 # CHECK-NEXT: [2]: Latency
  26 # CHECK-NEXT: [3]: RThroughput
  27 # CHECK-NEXT: [4]: MayLoad
  28 # CHECK-NEXT: [5]: MayStore
  29 # CHECK-NEXT: [6]: HasSideEffects (U)
  30
  31 # CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
  32 # CHECK-NEXT:  1      3     1.00                        vpopcntb        %zmm1, %zmm0
  33 # CHECK-NEXT:  2      11    1.00    *                   vpopcntb        (%rdi), %zmm0
  34 # CHECK-NEXT:  1      5     1.00                        vpopcntb        %zmm1, %zmm0 {%k1}
  35 # CHECK-NEXT:  2      13    1.00    *                   vpopcntb        (%rdi), %zmm0 {%k1}
  36 # CHECK-NEXT:  1      5     1.00                        vpopcntb        %zmm1, %zmm0 {%k1} {z}
  37 # CHECK-NEXT:  2      13    1.00    *                   vpopcntb        (%rdi), %zmm0 {%k1} {z}
  38 # CHECK-NEXT:  1      3     1.00                        vpopcntw        %zmm1, %zmm0
  39 # CHECK-NEXT:  2      11    1.00    *                   vpopcntw        (%rdi), %zmm0
  40 # CHECK-NEXT:  1      5     1.00                        vpopcntw        %zmm1, %zmm0 {%k1}
  41 # CHECK-NEXT:  2      13    1.00    *                   vpopcntw        (%rdi), %zmm0 {%k1}
  42 # CHECK-NEXT:  1      5     1.00                        vpopcntw        %zmm1, %zmm0 {%k1} {z}
  43 # CHECK-NEXT:  2      13    1.00    *                   vpopcntw        (%rdi), %zmm0 {%k1} {z}
  44 # CHECK-NEXT:  2      6     1.00                        vpshufbitqmb    %zmm16, %zmm17, %k2
  45 # CHECK-NEXT:  3      6     1.00    *                   vpshufbitqmb    (%rdi), %zmm17, %k2
  46 # CHECK-NEXT:  2      8     1.00                        vpshufbitqmb    %zmm16, %zmm17, %k2 {%k1}
  47 # CHECK-NEXT:  3      8     1.00    *                   vpshufbitqmb    (%rdi), %zmm17, %k2 {%k1}
  48
  49 # CHECK:      Resources:
  50 # CHECK-NEXT: [0]   - SPRPort00
  51 # CHECK-NEXT: [1]   - SPRPort01
  52 # CHECK-NEXT: [2]   - SPRPort02
  53 # CHECK-NEXT: [3]   - SPRPort03
  54 # CHECK-NEXT: [4]   - SPRPort04
  55 # CHECK-NEXT: [5]   - SPRPort05
  56 # CHECK-NEXT: [6]   - SPRPort06
  57 # CHECK-NEXT: [7]   - SPRPort07
  58 # CHECK-NEXT: [8]   - SPRPort08
  59 # CHECK-NEXT: [9]   - SPRPort09
  60 # CHECK-NEXT: [10]  - SPRPort10
  61 # CHECK-NEXT: [11]  - SPRPort11
  62 # CHECK-NEXT: [12]  - SPRPortInvalid
  63
  64 # CHECK:      Resource pressure per iteration:
  65 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]
  66 # CHECK-NEXT: 4.00    -     2.67   2.67    -     16.00   -      -      -      -      -     2.67    -
  67
  68 # CHECK:      Resource pressure by instruction:
  69 # CHECK-NEXT: [0]    [1]    [2]    [3]    [4]    [5]    [6]    [7]    [8]    [9]    [10]   [11]   [12]   Instructions:
  70 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntb       %zmm1, %zmm0
  71 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntb       (%rdi), %zmm0
  72 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntb       %zmm1, %zmm0 {%k1}
  73 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntb       (%rdi), %zmm0 {%k1}
  74 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntb       %zmm1, %zmm0 {%k1} {z}
  75 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntb       (%rdi), %zmm0 {%k1} {z}
  76 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntw       %zmm1, %zmm0
  77 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntw       (%rdi), %zmm0
  78 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntw       %zmm1, %zmm0 {%k1}
  79 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntw       (%rdi), %zmm0 {%k1}
  80 # CHECK-NEXT:  -      -      -      -      -     1.00    -      -      -      -      -      -      -     vpopcntw       %zmm1, %zmm0 {%k1} {z}
  81 # CHECK-NEXT:  -      -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpopcntw       (%rdi), %zmm0 {%k1} {z}
  82 # CHECK-NEXT: 1.00    -      -      -      -     1.00    -      -      -      -      -      -      -     vpshufbitqmb   %zmm16, %zmm17, %k2
  83 # CHECK-NEXT: 1.00    -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpshufbitqmb   (%rdi), %zmm17, %k2
  84 # CHECK-NEXT: 1.00    -      -      -      -     1.00    -      -      -      -      -      -      -     vpshufbitqmb   %zmm16, %zmm17, %k2 {%k1}
  85 # CHECK-NEXT: 1.00    -     0.33   0.33    -     1.00    -      -      -      -      -     0.33    -     vpshufbitqmb   (%rdi), %zmm17, %k2 {%k1}