|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5 |
| 2 | +; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=aarch64-unknown-linux \ |
| 3 | +; RUN: -mattr=-fullfp16 | FileCheck %s --check-prefixes=CHECK,NOFP16 |
| 4 | +; RUN: opt < %s -passes="print<cost-model>" 2>&1 -disable-output -mtriple=aarch64-unknown-linux \ |
| 5 | +; RUN: -mattr=+fullfp16 | FileCheck %s --check-prefixes=CHECK,FULLFP16 |
| 6 | + |
| 7 | +; res = lane 0 * lane 1 |
| 8 | +define double @extract_case1(<2 x double> %a) { |
| 9 | +; CHECK-LABEL: 'extract_case1' |
| 10 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 11 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x double> %a, i32 1 |
| 12 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %1 |
| 13 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 14 | +entry: |
| 15 | + %1 = extractelement <2 x double> %a, i32 0 |
| 16 | + %2 = extractelement <2 x double> %a, i32 1 |
| 17 | + %res = fmul double %1, %2 |
| 18 | + ret double %res |
| 19 | +} |
| 20 | + |
| 21 | +; res = lane 1 * lane 1 |
| 22 | +define double @extract_case2(<2 x double> %a) { |
| 23 | +; CHECK-LABEL: 'extract_case2' |
| 24 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %0 = extractelement <2 x double> %a, i32 1 |
| 25 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %0 |
| 26 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 27 | +entry: |
| 28 | + %1 = extractelement <2 x double> %a, i32 1 |
| 29 | + %res = fmul double %1, %1 |
| 30 | + ret double %res |
| 31 | +} |
| 32 | + |
| 33 | +; res = lane 0 * lane 0 |
| 34 | +define double @extract_case3(<2 x double> %a) { |
| 35 | +; CHECK-LABEL: 'extract_case3' |
| 36 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 37 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %0 |
| 38 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 39 | +entry: |
| 40 | + %1 = extractelement <2 x double> %a, i32 0 |
| 41 | + %res = fmul double %1, %1 |
| 42 | + ret double %res |
| 43 | +} |
| 44 | + |
| 45 | +; res = lane 0 * scalar |
| 46 | +define double @extract_case4(<2 x double> %a, double %b) { |
| 47 | +; CHECK-LABEL: 'extract_case4' |
| 48 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 49 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %b |
| 50 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 51 | +entry: |
| 52 | + %1 = extractelement <2 x double> %a, i32 0 |
| 53 | + %res = fmul double %1, %b |
| 54 | + ret double %res |
| 55 | +} |
| 56 | + |
| 57 | +; res = lane 1 * scalar |
| 58 | +define double @extract_case5(<2 x double> %a, double %b) { |
| 59 | +; CHECK-LABEL: 'extract_case5' |
| 60 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %0 = extractelement <2 x double> %a, i32 1 |
| 61 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %b |
| 62 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 63 | +entry: |
| 64 | + %1 = extractelement <2 x double> %a, i32 1 |
| 65 | + %res = fmul double %1, %b |
| 66 | + ret double %res |
| 67 | +} |
| 68 | + |
| 69 | +; Input vector = <3 x double> (i.e. odd length vector) |
| 70 | +; res = lane 0 * lane 1 |
| 71 | +define double @extract_case6(<3 x double> %a) { |
| 72 | +; CHECK-LABEL: 'extract_case6' |
| 73 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <3 x double> %a, i32 0 |
| 74 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <3 x double> %a, i32 1 |
| 75 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %1 |
| 76 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 77 | +entry: |
| 78 | + %1 = extractelement <3 x double> %a, i32 0 |
| 79 | + %2 = extractelement <3 x double> %a, i32 1 |
| 80 | + %res = fmul double %1, %2 |
| 81 | + ret double %res |
| 82 | +} |
| 83 | + |
| 84 | +; res = lane 1 * lane 2 |
| 85 | +; Extract from lane 2 is equivalent to extract from lane 0 of other 128-bit |
| 86 | +; register. But for other register sizes, this is not the case. |
| 87 | +define double @extract_case7(<4 x double> %a) { |
| 88 | +; CHECK-LABEL: 'extract_case7' |
| 89 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %0 = extractelement <4 x double> %a, i32 1 |
| 90 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %1 = extractelement <4 x double> %a, i32 2 |
| 91 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul double %0, %1 |
| 92 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 93 | +entry: |
| 94 | + %1 = extractelement <4 x double> %a, i32 1 |
| 95 | + %2 = extractelement <4 x double> %a, i32 2 |
| 96 | + %res = fmul double %1, %2 |
| 97 | + ret double %res |
| 98 | +} |
| 99 | + |
| 100 | +; res = lane 0 * lane 1 |
| 101 | +; Additional insert of extract from lane 1. |
| 102 | +define double @extract_case8(<2 x double> %a) { |
| 103 | +; CHECK-LABEL: 'extract_case8' |
| 104 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 105 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x double> %a, i32 1 |
| 106 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %2 = insertelement <2 x double> %a, double %1, i32 0 |
| 107 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 8 for instruction: %3 = call double @llvm.vector.reduce.fmul.v2f64(double 0.000000e+00, <2 x double> %2) |
| 108 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %4 = fmul double %0, %1 |
| 109 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %5 = fmul double %3, %4 |
| 110 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %5 |
| 111 | +entry: |
| 112 | + %1 = extractelement <2 x double> %a, i32 0 |
| 113 | + %2 = extractelement <2 x double> %a, i32 1 |
| 114 | + %3 = insertelement <2 x double> %a, double %2, i32 0 |
| 115 | + %4 = call double @llvm.vector.reduce.fmul.v2f64(double 0.0, <2 x double> %3) |
| 116 | + %5 = fmul double %1, %2 |
| 117 | + %6 = fmul double %4, %5 |
| 118 | + ret double %6 |
| 119 | +} |
| 120 | + |
| 121 | +; res = lane 0 * lane 1 |
| 122 | +; Additional insert of extract from lane 1. |
| 123 | +define double @extract_case9(<2 x double> %a) { |
| 124 | +; CHECK-LABEL: 'extract_case9' |
| 125 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 126 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x double> %a, i32 1 |
| 127 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %2 = insertelement <2 x double> %a, double %1, i32 0 |
| 128 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 6 for instruction: %3 = call double @llvm.vector.reduce.fadd.v2f64(double 0.000000e+00, <2 x double> %2) |
| 129 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %4 = fmul double %0, %1 |
| 130 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %5 = fmul double %3, %4 |
| 131 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %5 |
| 132 | +entry: |
| 133 | + %1 = extractelement <2 x double> %a, i32 0 |
| 134 | + %2 = extractelement <2 x double> %a, i32 1 |
| 135 | + %3 = insertelement <2 x double> %a, double %2, i32 0 |
| 136 | + %4 = call double @llvm.vector.reduce.fadd.v2f64(double 0.0, <2 x double> %3) |
| 137 | + %5 = fmul double %1, %2 |
| 138 | + %6 = fmul double %4, %5 |
| 139 | + ret double %6 |
| 140 | +} |
| 141 | + |
| 142 | +; res = lane 0 * lane 1 |
| 143 | +; Extract from lane 1 passed as function param. |
| 144 | +define double @extract_case10(<4 x double> %a) { |
| 145 | +; CHECK-LABEL: 'extract_case10' |
| 146 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <4 x double> %a, i32 0 |
| 147 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <4 x double> %a, i32 1 |
| 148 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: call void @foo(double %1) |
| 149 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %2 = fmul double %0, %1 |
| 150 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %2 |
| 151 | +entry: |
| 152 | + %1 = extractelement <4 x double> %a, i32 0 |
| 153 | + %2 = extractelement <4 x double> %a, i32 1 |
| 154 | + call void @foo(double %2) |
| 155 | + %3 = fmul double %1, %2 |
| 156 | + ret double %3 |
| 157 | +} |
| 158 | + |
| 159 | +; res = lane 0 * lane 1 |
| 160 | +define half @extract_case11(<2 x half> %a) { |
| 161 | +; CHECK-LABEL: 'extract_case11' |
| 162 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x half> %a, i32 0 |
| 163 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x half> %a, i32 1 |
| 164 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul half %0, %1 |
| 165 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret half %res |
| 166 | +entry: |
| 167 | + %1 = extractelement <2 x half> %a, i32 0 |
| 168 | + %2 = extractelement <2 x half> %a, i32 1 |
| 169 | + %res = fmul half %1, %2 |
| 170 | + ret half %res |
| 171 | +} |
| 172 | + |
| 173 | +; res = lane 0 * lane 1 |
| 174 | +define float @extract_case12(<2 x float> %a) { |
| 175 | +; CHECK-LABEL: 'extract_case12' |
| 176 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x float> %a, i32 0 |
| 177 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x float> %a, i32 1 |
| 178 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %res = fmul float %0, %1 |
| 179 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret float %res |
| 180 | +entry: |
| 181 | + %1 = extractelement <2 x float> %a, i32 0 |
| 182 | + %2 = extractelement <2 x float> %a, i32 1 |
| 183 | + %res = fmul float %1, %2 |
| 184 | + ret float %res |
| 185 | +} |
| 186 | + |
| 187 | +; res = lane 0 + lane 1 |
| 188 | +; Use of bin-op other than fmul. |
| 189 | +define double @extract_case13(<2 x double> %a) { |
| 190 | +; CHECK-LABEL: 'extract_case13' |
| 191 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: %0 = extractelement <2 x double> %a, i32 0 |
| 192 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 2 for instruction: %1 = extractelement <2 x double> %a, i32 1 |
| 193 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 1 for instruction: %res = fadd double %0, %1 |
| 194 | +; CHECK-NEXT: Cost Model: Found an estimated cost of 0 for instruction: ret double %res |
| 195 | +entry: |
| 196 | + %1 = extractelement <2 x double> %a, i32 0 |
| 197 | + %2 = extractelement <2 x double> %a, i32 1 |
| 198 | + %res = fadd double %1, %2 |
| 199 | + ret double %res |
| 200 | +} |
| 201 | + |
| 202 | +declare void @foo(double) |
| 203 | +;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: |
| 204 | +; FULLFP16: {{.*}} |
| 205 | +; NOFP16: {{.*}} |
0 commit comments