|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4 |
| 2 | +; RUN: llc < %s -march=bpf | FileCheck %s |
| 3 | + |
| 4 | +; test that we can expand CTTZ & CTLZ |
| 5 | + |
| 6 | +declare i32 @llvm.cttz.i32(i32, i1) |
| 7 | + |
| 8 | +define i32 @cttz_i32_zdef(i32 %a) { |
| 9 | +; CHECK-LABEL: cttz_i32_zdef: |
| 10 | +; CHECK: # %bb.0: |
| 11 | +; CHECK-NEXT: r2 = r1 |
| 12 | +; CHECK-NEXT: r2 = -r2 |
| 13 | +; CHECK-NEXT: r1 &= r2 |
| 14 | +; CHECK-NEXT: r1 *= 125613361 |
| 15 | +; CHECK-NEXT: r2 = 4160749568 ll |
| 16 | +; CHECK-NEXT: r1 &= r2 |
| 17 | +; CHECK-NEXT: r1 >>= 27 |
| 18 | +; CHECK-NEXT: r2 = {{\.?LCPI[0-9]+_[0-9]+}} ll |
| 19 | +; CHECK-NEXT: r2 += r1 |
| 20 | +; CHECK-NEXT: r0 = *(u8 *)(r2 + 0) |
| 21 | +; CHECK-NEXT: exit |
| 22 | + %ret = call i32 @llvm.cttz.i32(i32 %a, i1 1) |
| 23 | + ret i32 %ret |
| 24 | +} |
| 25 | + |
| 26 | + |
| 27 | +define i32 @cttz_i32(i32 %a) { |
| 28 | +; CHECK-LABEL: cttz_i32: |
| 29 | +; CHECK: # %bb.0: |
| 30 | +; CHECK-NEXT: r0 = 32 |
| 31 | +; CHECK-NEXT: r2 = r1 |
| 32 | +; CHECK-NEXT: r2 <<= 32 |
| 33 | +; CHECK-NEXT: r2 >>= 32 |
| 34 | +; CHECK-NEXT: if r2 == 0 goto LBB1_2 |
| 35 | +; CHECK-NEXT: # %bb.1: # %cond.false |
| 36 | +; CHECK-NEXT: r2 = r1 |
| 37 | +; CHECK-NEXT: r2 = -r2 |
| 38 | +; CHECK-NEXT: r1 &= r2 |
| 39 | +; CHECK-NEXT: r1 *= 125613361 |
| 40 | +; CHECK-NEXT: r2 = 4160749568 ll |
| 41 | +; CHECK-NEXT: r1 &= r2 |
| 42 | +; CHECK-NEXT: r1 >>= 27 |
| 43 | +; CHECK-NEXT: r2 = {{\.?LCPI[0-9]+_[0-9]+}} ll |
| 44 | +; CHECK-NEXT: r2 += r1 |
| 45 | +; CHECK-NEXT: r0 = *(u8 *)(r2 + 0) |
| 46 | +; CHECK-NEXT: LBB1_2: # %cond.end |
| 47 | +; CHECK-NEXT: exit |
| 48 | + %ret = call i32 @llvm.cttz.i32(i32 %a, i1 0) |
| 49 | + ret i32 %ret |
| 50 | +} |
| 51 | + |
| 52 | +declare i64 @llvm.cttz.i64(i64, i1) |
| 53 | + |
| 54 | +define i64 @cttz_i64_zdef(i64 %a) { |
| 55 | +; CHECK-LABEL: cttz_i64_zdef: |
| 56 | +; CHECK: # %bb.0: |
| 57 | +; CHECK-NEXT: r2 = r1 |
| 58 | +; CHECK-NEXT: r2 = -r2 |
| 59 | +; CHECK-NEXT: r1 &= r2 |
| 60 | +; CHECK-NEXT: r2 = 151050438420815295 ll |
| 61 | +; CHECK-NEXT: r1 *= r2 |
| 62 | +; CHECK-NEXT: r1 >>= 58 |
| 63 | +; CHECK-NEXT: r2 = {{\.?LCPI[0-9]+_[0-9]+}} ll |
| 64 | +; CHECK-NEXT: r2 += r1 |
| 65 | +; CHECK-NEXT: r0 = *(u8 *)(r2 + 0) |
| 66 | +; CHECK-NEXT: exit |
| 67 | + %ret = call i64 @llvm.cttz.i64(i64 %a, i1 1) |
| 68 | + ret i64 %ret |
| 69 | +} |
| 70 | + |
| 71 | + |
| 72 | +define i64 @cttz_i64(i64 %a) { |
| 73 | +; CHECK-LABEL: cttz_i64: |
| 74 | +; CHECK: # %bb.0: |
| 75 | +; CHECK-NEXT: r0 = 64 |
| 76 | +; CHECK-NEXT: if r1 == 0 goto LBB3_2 |
| 77 | +; CHECK-NEXT: # %bb.1: # %cond.false |
| 78 | +; CHECK-NEXT: r2 = r1 |
| 79 | +; CHECK-NEXT: r2 = -r2 |
| 80 | +; CHECK-NEXT: r1 &= r2 |
| 81 | +; CHECK-NEXT: r2 = 151050438420815295 ll |
| 82 | +; CHECK-NEXT: r1 *= r2 |
| 83 | +; CHECK-NEXT: r1 >>= 58 |
| 84 | +; CHECK-NEXT: r2 = {{\.?LCPI[0-9]+_[0-9]+}} ll |
| 85 | +; CHECK-NEXT: r2 += r1 |
| 86 | +; CHECK-NEXT: r0 = *(u8 *)(r2 + 0) |
| 87 | +; CHECK-NEXT: LBB3_2: # %cond.end |
| 88 | +; CHECK-NEXT: exit |
| 89 | + %ret = call i64 @llvm.cttz.i64(i64 %a, i1 0) |
| 90 | + ret i64 %ret |
| 91 | +} |
| 92 | + |
| 93 | + |
| 94 | +declare i32 @llvm.ctlz.i32(i32, i1) |
| 95 | + |
| 96 | +define i32 @ctlz_i32_zdef(i32 %a) { |
| 97 | +; CHECK-LABEL: ctlz_i32_zdef: |
| 98 | +; CHECK: # %bb.0: |
| 99 | +; CHECK-NEXT: r2 = 4294967294 ll |
| 100 | +; CHECK-NEXT: r3 = r1 |
| 101 | +; CHECK-NEXT: r3 &= r2 |
| 102 | +; CHECK-NEXT: r3 >>= 1 |
| 103 | +; CHECK-NEXT: r1 |= r3 |
| 104 | +; CHECK-NEXT: r2 = 4294967292 ll |
| 105 | +; CHECK-NEXT: r3 = r1 |
| 106 | +; CHECK-NEXT: r3 &= r2 |
| 107 | +; CHECK-NEXT: r3 >>= 2 |
| 108 | +; CHECK-NEXT: r1 |= r3 |
| 109 | +; CHECK-NEXT: r2 = 4294967280 ll |
| 110 | +; CHECK-NEXT: r3 = r1 |
| 111 | +; CHECK-NEXT: r3 &= r2 |
| 112 | +; CHECK-NEXT: r3 >>= 4 |
| 113 | +; CHECK-NEXT: r1 |= r3 |
| 114 | +; CHECK-NEXT: r2 = 4294967040 ll |
| 115 | +; CHECK-NEXT: r3 = r1 |
| 116 | +; CHECK-NEXT: r3 &= r2 |
| 117 | +; CHECK-NEXT: r3 >>= 8 |
| 118 | +; CHECK-NEXT: r1 |= r3 |
| 119 | +; CHECK-NEXT: r2 = 4294901760 ll |
| 120 | +; CHECK-NEXT: r3 = r1 |
| 121 | +; CHECK-NEXT: r3 &= r2 |
| 122 | +; CHECK-NEXT: r3 >>= 16 |
| 123 | +; CHECK-NEXT: r1 |= r3 |
| 124 | +; CHECK-NEXT: r1 ^= -1 |
| 125 | +; CHECK-NEXT: r2 = r1 |
| 126 | +; CHECK-NEXT: r2 >>= 1 |
| 127 | +; CHECK-NEXT: r2 &= 1431655765 |
| 128 | +; CHECK-NEXT: r1 -= r2 |
| 129 | +; CHECK-NEXT: r0 = r1 |
| 130 | +; CHECK-NEXT: r0 &= 858993459 |
| 131 | +; CHECK-NEXT: r1 >>= 2 |
| 132 | +; CHECK-NEXT: r1 &= 858993459 |
| 133 | +; CHECK-NEXT: r0 += r1 |
| 134 | +; CHECK-NEXT: r1 = r0 |
| 135 | +; CHECK-NEXT: r1 >>= 4 |
| 136 | +; CHECK-NEXT: r0 += r1 |
| 137 | +; CHECK-NEXT: r0 &= 252645135 |
| 138 | +; CHECK-NEXT: r0 *= 16843009 |
| 139 | +; CHECK-NEXT: r1 = 4278190080 ll |
| 140 | +; CHECK-NEXT: r0 &= r1 |
| 141 | +; CHECK-NEXT: r0 >>= 24 |
| 142 | +; CHECK-NEXT: exit |
| 143 | + %ret = call i32 @llvm.ctlz.i32(i32 %a, i1 1) |
| 144 | + ret i32 %ret |
| 145 | +} |
| 146 | + |
| 147 | + |
| 148 | +define i32 @ctlz_i32(i32 %a) { |
| 149 | +; CHECK-LABEL: ctlz_i32: |
| 150 | +; CHECK: # %bb.0: |
| 151 | +; CHECK-NEXT: r0 = 32 |
| 152 | +; CHECK-NEXT: r2 = r1 |
| 153 | +; CHECK-NEXT: r2 <<= 32 |
| 154 | +; CHECK-NEXT: r2 >>= 32 |
| 155 | +; CHECK-NEXT: if r2 == 0 goto LBB5_2 |
| 156 | +; CHECK-NEXT: # %bb.1: # %cond.false |
| 157 | +; CHECK-NEXT: r2 = 4294967294 ll |
| 158 | +; CHECK-NEXT: r3 = r1 |
| 159 | +; CHECK-NEXT: r3 &= r2 |
| 160 | +; CHECK-NEXT: r3 >>= 1 |
| 161 | +; CHECK-NEXT: r1 |= r3 |
| 162 | +; CHECK-NEXT: r2 = 4294967292 ll |
| 163 | +; CHECK-NEXT: r3 = r1 |
| 164 | +; CHECK-NEXT: r3 &= r2 |
| 165 | +; CHECK-NEXT: r3 >>= 2 |
| 166 | +; CHECK-NEXT: r1 |= r3 |
| 167 | +; CHECK-NEXT: r2 = 4294967280 ll |
| 168 | +; CHECK-NEXT: r3 = r1 |
| 169 | +; CHECK-NEXT: r3 &= r2 |
| 170 | +; CHECK-NEXT: r3 >>= 4 |
| 171 | +; CHECK-NEXT: r1 |= r3 |
| 172 | +; CHECK-NEXT: r2 = 4294967040 ll |
| 173 | +; CHECK-NEXT: r3 = r1 |
| 174 | +; CHECK-NEXT: r3 &= r2 |
| 175 | +; CHECK-NEXT: r3 >>= 8 |
| 176 | +; CHECK-NEXT: r1 |= r3 |
| 177 | +; CHECK-NEXT: r2 = 4294901760 ll |
| 178 | +; CHECK-NEXT: r3 = r1 |
| 179 | +; CHECK-NEXT: r3 &= r2 |
| 180 | +; CHECK-NEXT: r3 >>= 16 |
| 181 | +; CHECK-NEXT: r1 |= r3 |
| 182 | +; CHECK-NEXT: r1 ^= -1 |
| 183 | +; CHECK-NEXT: r2 = r1 |
| 184 | +; CHECK-NEXT: r2 >>= 1 |
| 185 | +; CHECK-NEXT: r2 &= 1431655765 |
| 186 | +; CHECK-NEXT: r1 -= r2 |
| 187 | +; CHECK-NEXT: r0 = r1 |
| 188 | +; CHECK-NEXT: r0 &= 858993459 |
| 189 | +; CHECK-NEXT: r1 >>= 2 |
| 190 | +; CHECK-NEXT: r1 &= 858993459 |
| 191 | +; CHECK-NEXT: r0 += r1 |
| 192 | +; CHECK-NEXT: r1 = r0 |
| 193 | +; CHECK-NEXT: r1 >>= 4 |
| 194 | +; CHECK-NEXT: r0 += r1 |
| 195 | +; CHECK-NEXT: r0 &= 252645135 |
| 196 | +; CHECK-NEXT: r0 *= 16843009 |
| 197 | +; CHECK-NEXT: r1 = 4278190080 ll |
| 198 | +; CHECK-NEXT: r0 &= r1 |
| 199 | +; CHECK-NEXT: r0 >>= 24 |
| 200 | +; CHECK-NEXT: LBB5_2: # %cond.end |
| 201 | +; CHECK-NEXT: exit |
| 202 | + %ret = call i32 @llvm.ctlz.i32(i32 %a, i1 0) |
| 203 | + ret i32 %ret |
| 204 | +} |
| 205 | + |
| 206 | +declare i64 @llvm.ctlz.i64(i64, i1) |
| 207 | + |
| 208 | +define i64 @ctlz_i64_zdef(i64 %a) { |
| 209 | +; CHECK-LABEL: ctlz_i64_zdef: |
| 210 | +; CHECK: # %bb.0: |
| 211 | +; CHECK-NEXT: r2 = r1 |
| 212 | +; CHECK-NEXT: r2 >>= 1 |
| 213 | +; CHECK-NEXT: r1 |= r2 |
| 214 | +; CHECK-NEXT: r2 = r1 |
| 215 | +; CHECK-NEXT: r2 >>= 2 |
| 216 | +; CHECK-NEXT: r1 |= r2 |
| 217 | +; CHECK-NEXT: r2 = r1 |
| 218 | +; CHECK-NEXT: r2 >>= 4 |
| 219 | +; CHECK-NEXT: r1 |= r2 |
| 220 | +; CHECK-NEXT: r2 = r1 |
| 221 | +; CHECK-NEXT: r2 >>= 8 |
| 222 | +; CHECK-NEXT: r1 |= r2 |
| 223 | +; CHECK-NEXT: r2 = r1 |
| 224 | +; CHECK-NEXT: r2 >>= 16 |
| 225 | +; CHECK-NEXT: r1 |= r2 |
| 226 | +; CHECK-NEXT: r2 = r1 |
| 227 | +; CHECK-NEXT: r2 >>= 32 |
| 228 | +; CHECK-NEXT: r1 |= r2 |
| 229 | +; CHECK-NEXT: r1 ^= -1 |
| 230 | +; CHECK-NEXT: r2 = 6148914691236517205 ll |
| 231 | +; CHECK-NEXT: r3 = r1 |
| 232 | +; CHECK-NEXT: r3 >>= 1 |
| 233 | +; CHECK-NEXT: r3 &= r2 |
| 234 | +; CHECK-NEXT: r1 -= r3 |
| 235 | +; CHECK-NEXT: r2 = 3689348814741910323 ll |
| 236 | +; CHECK-NEXT: r0 = r1 |
| 237 | +; CHECK-NEXT: r0 &= r2 |
| 238 | +; CHECK-NEXT: r1 >>= 2 |
| 239 | +; CHECK-NEXT: r1 &= r2 |
| 240 | +; CHECK-NEXT: r0 += r1 |
| 241 | +; CHECK-NEXT: r1 = r0 |
| 242 | +; CHECK-NEXT: r1 >>= 4 |
| 243 | +; CHECK-NEXT: r0 += r1 |
| 244 | +; CHECK-NEXT: r1 = 1085102592571150095 ll |
| 245 | +; CHECK-NEXT: r0 &= r1 |
| 246 | +; CHECK-NEXT: r1 = 72340172838076673 ll |
| 247 | +; CHECK-NEXT: r0 *= r1 |
| 248 | +; CHECK-NEXT: r0 >>= 56 |
| 249 | +; CHECK-NEXT: exit |
| 250 | + %ret = call i64 @llvm.ctlz.i64(i64 %a, i1 1) |
| 251 | + ret i64 %ret |
| 252 | +} |
| 253 | + |
| 254 | + |
| 255 | +define i64 @ctlz_i64(i64 %a) { |
| 256 | +; CHECK-LABEL: ctlz_i64: |
| 257 | +; CHECK: # %bb.0: |
| 258 | +; CHECK-NEXT: r0 = 64 |
| 259 | +; CHECK-NEXT: if r1 == 0 goto LBB7_2 |
| 260 | +; CHECK-NEXT: # %bb.1: # %cond.false |
| 261 | +; CHECK-NEXT: r2 = r1 |
| 262 | +; CHECK-NEXT: r2 >>= 1 |
| 263 | +; CHECK-NEXT: r1 |= r2 |
| 264 | +; CHECK-NEXT: r2 = r1 |
| 265 | +; CHECK-NEXT: r2 >>= 2 |
| 266 | +; CHECK-NEXT: r1 |= r2 |
| 267 | +; CHECK-NEXT: r2 = r1 |
| 268 | +; CHECK-NEXT: r2 >>= 4 |
| 269 | +; CHECK-NEXT: r1 |= r2 |
| 270 | +; CHECK-NEXT: r2 = r1 |
| 271 | +; CHECK-NEXT: r2 >>= 8 |
| 272 | +; CHECK-NEXT: r1 |= r2 |
| 273 | +; CHECK-NEXT: r2 = r1 |
| 274 | +; CHECK-NEXT: r2 >>= 16 |
| 275 | +; CHECK-NEXT: r1 |= r2 |
| 276 | +; CHECK-NEXT: r2 = r1 |
| 277 | +; CHECK-NEXT: r2 >>= 32 |
| 278 | +; CHECK-NEXT: r1 |= r2 |
| 279 | +; CHECK-NEXT: r1 ^= -1 |
| 280 | +; CHECK-NEXT: r2 = 6148914691236517205 ll |
| 281 | +; CHECK-NEXT: r3 = r1 |
| 282 | +; CHECK-NEXT: r3 >>= 1 |
| 283 | +; CHECK-NEXT: r3 &= r2 |
| 284 | +; CHECK-NEXT: r1 -= r3 |
| 285 | +; CHECK-NEXT: r2 = 3689348814741910323 ll |
| 286 | +; CHECK-NEXT: r0 = r1 |
| 287 | +; CHECK-NEXT: r0 &= r2 |
| 288 | +; CHECK-NEXT: r1 >>= 2 |
| 289 | +; CHECK-NEXT: r1 &= r2 |
| 290 | +; CHECK-NEXT: r0 += r1 |
| 291 | +; CHECK-NEXT: r1 = r0 |
| 292 | +; CHECK-NEXT: r1 >>= 4 |
| 293 | +; CHECK-NEXT: r0 += r1 |
| 294 | +; CHECK-NEXT: r1 = 1085102592571150095 ll |
| 295 | +; CHECK-NEXT: r0 &= r1 |
| 296 | +; CHECK-NEXT: r1 = 72340172838076673 ll |
| 297 | +; CHECK-NEXT: r0 *= r1 |
| 298 | +; CHECK-NEXT: r0 >>= 56 |
| 299 | +; CHECK-NEXT: LBB7_2: # %cond.end |
| 300 | +; CHECK-NEXT: exit |
| 301 | + %ret = call i64 @llvm.ctlz.i64(i64 %a, i1 0) |
| 302 | + ret i64 %ret |
| 303 | +} |
| 304 | + |
0 commit comments