Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 1 | ; RUN: opt -analyze -scalar-evolution < %s | FileCheck %s |
| 2 | |
| 3 | ; This test set ensures that we can correctly operate with recurrencies from |
| 4 | ; different loops. |
| 5 | |
| 6 | ; Check that we can evaluate a sum of phis from two different loops in any |
| 7 | ; order. |
| 8 | |
| 9 | define void @test_00() { |
| 10 | |
| 11 | ; CHECK-LABEL: Classifying expressions for: @test_00 |
| 12 | ; CHECK: %sum1 = add i32 %phi1, %phi2 |
| 13 | ; CHECK-NEXT: --> {14,+,3}<%loop1> |
| 14 | ; CHECK: %sum2 = add i32 %sum1, %phi3 |
| 15 | ; CHECK-NEXT: --> {20,+,6}<%loop1> |
| 16 | ; CHECK: %sum3 = add i32 %phi4, %phi5 |
| 17 | ; CHECK-NEXT: --> {116,+,3}<%loop2> |
| 18 | ; CHECK: %sum4 = add i32 %sum3, %phi6 |
| 19 | ; CHECK-NEXT: --> {159,+,6}<%loop2> |
| 20 | ; CHECK: %s1 = add i32 %phi1, %phi4 |
| 21 | ; CHECK-NEXT: --> {{{{}}73,+,1}<%loop1>,+,1}<%loop2> |
| 22 | ; CHECK: %s2 = add i32 %phi5, %phi2 |
| 23 | ; CHECK-NEXT: --> {{{{}}57,+,2}<%loop1>,+,2}<%loop2> |
| 24 | ; CHECK: %s3 = add i32 %sum1, %sum3 |
| 25 | ; CHECK-NEXT: --> {{{{}}130,+,3}<%loop1>,+,3}<%loop2> |
| 26 | ; CHECK: %s4 = add i32 %sum4, %sum2 |
| 27 | ; CHECK-NEXT: --> {{{{}}179,+,6}<%loop1>,+,6}<%loop2> |
| 28 | ; CHECK: %s5 = add i32 %phi3, %sum3 |
| 29 | ; CHECK-NEXT: --> {{{{}}122,+,3}<%loop1>,+,3}<%loop2> |
| 30 | ; CHECK: %s6 = add i32 %sum2, %phi6 |
| 31 | ; CHECK-NEXT: --> {{{{}}63,+,6}<%loop1>,+,3}<%loop2> |
| 32 | |
| 33 | entry: |
| 34 | br label %loop1 |
| 35 | |
| 36 | loop1: |
| 37 | %phi1 = phi i32 [ 10, %entry ], [ %phi1.inc, %loop1 ] |
| 38 | %phi2 = phi i32 [ 4, %entry ], [ %phi2.inc, %loop1 ] |
| 39 | %phi3 = phi i32 [ 6, %entry ], [ %phi3.inc, %loop1 ] |
| 40 | %phi1.inc = add i32 %phi1, 1 |
| 41 | %phi2.inc = add i32 %phi2, 2 |
| 42 | %phi3.inc = add i32 %phi3, 3 |
| 43 | %sum1 = add i32 %phi1, %phi2 |
| 44 | %sum2 = add i32 %sum1, %phi3 |
| 45 | %cond1 = icmp ult i32 %sum2, 1000 |
| 46 | br i1 %cond1, label %loop1, label %loop2 |
| 47 | |
| 48 | loop2: |
| 49 | %phi4 = phi i32 [ 63, %loop1 ], [ %phi4.inc, %loop2 ] |
| 50 | %phi5 = phi i32 [ 53, %loop1 ], [ %phi5.inc, %loop2 ] |
| 51 | %phi6 = phi i32 [ 43, %loop1 ], [ %phi6.inc, %loop2 ] |
| 52 | %phi4.inc = add i32 %phi4, 1 |
| 53 | %phi5.inc = add i32 %phi5, 2 |
| 54 | %phi6.inc = add i32 %phi6, 3 |
| 55 | %sum3 = add i32 %phi4, %phi5 |
| 56 | %sum4 = add i32 %sum3, %phi6 |
| 57 | %cond2 = icmp ult i32 %sum4, 1000 |
| 58 | br i1 %cond2, label %loop2, label %exit |
| 59 | |
| 60 | exit: |
| 61 | %s1 = add i32 %phi1, %phi4 |
| 62 | %s2 = add i32 %phi5, %phi2 |
| 63 | %s3 = add i32 %sum1, %sum3 |
| 64 | %s4 = add i32 %sum4, %sum2 |
| 65 | %s5 = add i32 %phi3, %sum3 |
| 66 | %s6 = add i32 %sum2, %phi6 |
| 67 | ret void |
| 68 | } |
| 69 | |
| 70 | ; Check that we can evaluate a sum of phis+invariants from two different loops |
| 71 | ; in any order. |
| 72 | |
| 73 | define void @test_01(i32 %a, i32 %b) { |
| 74 | |
| 75 | ; CHECK-LABEL: Classifying expressions for: @test_01 |
| 76 | ; CHECK: %sum1 = add i32 %phi1, %phi2 |
| 77 | ; CHECK-NEXT: --> {(%a + %b),+,3}<%loop1> |
| 78 | ; CHECK: %sum2 = add i32 %sum1, %phi3 |
| 79 | ; CHECK-NEXT: --> {(6 + %a + %b),+,6}<%loop1> |
| 80 | ; CHECK: %is1 = add i32 %sum2, %a |
| 81 | ; CHECK-NEXT: --> {(6 + (2 * %a) + %b),+,6}<%loop1> |
| 82 | ; CHECK: %sum3 = add i32 %phi4, %phi5 |
| 83 | ; CHECK-NEXT: --> {116,+,3}<%loop2> |
| 84 | ; CHECK: %sum4 = add i32 %sum3, %phi6 |
| 85 | ; CHECK-NEXT: --> {159,+,6}<%loop2> |
| 86 | ; CHECK: %is2 = add i32 %sum4, %b |
| 87 | ; CHECK-NEXT: --> {(159 + %b),+,6}<%loop2> |
| 88 | ; CHECK: %ec2 = add i32 %is1, %is2 |
| 89 | ; CHECK-NEXT: --> {{{{}}(165 + (2 * %a) + (2 * %b)),+,6}<%loop1>,+,6}<%loop2> |
| 90 | ; CHECK: %s1 = add i32 %phi1, %is1 |
| 91 | ; CHECK-NEXT: --> {(6 + (3 * %a) + %b),+,7}<%loop1> |
| 92 | ; CHECK: %s2 = add i32 %is2, %phi4 |
| 93 | ; CHECK-NEXT: --> {(222 + %b),+,7}<%loop2> |
| 94 | ; CHECK: %s3 = add i32 %is1, %phi5 |
| 95 | ; CHECK-NEXT: --> {{{{}}(59 + (2 * %a) + %b),+,6}<%loop1>,+,2}<%loop2> |
| 96 | ; CHECK: %s4 = add i32 %phi2, %is2 |
| 97 | ; CHECK-NEXT: --> {{{{}}(159 + (2 * %b)),+,2}<%loop1>,+,6}<%loop2> |
| 98 | ; CHECK: %s5 = add i32 %is1, %is2 |
| 99 | ; CHECK-NEXT: --> {{{{}}(165 + (2 * %a) + (2 * %b)),+,6}<%loop1>,+,6}<%loop2> |
| 100 | ; CHECK: %s6 = add i32 %is2, %is1 |
| 101 | ; CHECK-NEXT: --> {{{{}}(165 + (2 * %a) + (2 * %b)),+,6}<%loop1>,+,6}<%loop2> |
| 102 | |
| 103 | entry: |
| 104 | br label %loop1 |
| 105 | |
| 106 | loop1: |
| 107 | %phi1 = phi i32 [ %a, %entry ], [ %phi1.inc, %loop1 ] |
| 108 | %phi2 = phi i32 [ %b, %entry ], [ %phi2.inc, %loop1 ] |
| 109 | %phi3 = phi i32 [ 6, %entry ], [ %phi3.inc, %loop1 ] |
| 110 | %phi1.inc = add i32 %phi1, 1 |
| 111 | %phi2.inc = add i32 %phi2, 2 |
| 112 | %phi3.inc = add i32 %phi3, 3 |
| 113 | %sum1 = add i32 %phi1, %phi2 |
| 114 | %sum2 = add i32 %sum1, %phi3 |
| 115 | %is1 = add i32 %sum2, %a |
| 116 | %cond1 = icmp ult i32 %is1, 1000 |
| 117 | br i1 %cond1, label %loop1, label %loop2 |
| 118 | |
| 119 | loop2: |
| 120 | %phi4 = phi i32 [ 63, %loop1 ], [ %phi4.inc, %loop2 ] |
| 121 | %phi5 = phi i32 [ 53, %loop1 ], [ %phi5.inc, %loop2 ] |
| 122 | %phi6 = phi i32 [ 43, %loop1 ], [ %phi6.inc, %loop2 ] |
| 123 | %phi4.inc = add i32 %phi4, 1 |
| 124 | %phi5.inc = add i32 %phi5, 2 |
| 125 | %phi6.inc = add i32 %phi6, 3 |
| 126 | %sum3 = add i32 %phi4, %phi5 |
| 127 | %sum4 = add i32 %sum3, %phi6 |
| 128 | %is2 = add i32 %sum4, %b |
| 129 | %ec2 = add i32 %is1, %is2 |
| 130 | %cond2 = icmp ult i32 %ec2, 1000 |
| 131 | br i1 %cond2, label %loop2, label %exit |
| 132 | |
| 133 | exit: |
| 134 | %s1 = add i32 %phi1, %is1 |
| 135 | %s2 = add i32 %is2, %phi4 |
| 136 | %s3 = add i32 %is1, %phi5 |
| 137 | %s4 = add i32 %phi2, %is2 |
| 138 | %s5 = add i32 %is1, %is2 |
| 139 | %s6 = add i32 %is2, %is1 |
| 140 | ret void |
| 141 | } |
| 142 | |
| 143 | ; Check that we can correctly evaluate a sum of phis+variants from two different |
| 144 | ; loops in any order. |
| 145 | |
| 146 | define void @test_02(i32 %a, i32 %b, i32* %p) { |
| 147 | |
| 148 | ; CHECK-LABEL: Classifying expressions for: @test_02 |
| 149 | ; CHECK: %sum1 = add i32 %phi1, %phi2 |
| 150 | ; CHECK-NEXT: --> {(%a + %b),+,3}<%loop1> |
| 151 | ; CHECK: %sum2 = add i32 %sum1, %phi3 |
| 152 | ; CHECK-NEXT: --> {(6 + %a + %b),+,6}<%loop1> |
| 153 | ; CHECK: %is1 = add i32 %sum2, %v1 |
| 154 | ; CHECK-NEXT: --> ({(6 + %a + %b),+,6}<%loop1> + %v1) |
| 155 | ; CHECK: %sum3 = add i32 %phi4, %phi5 |
| 156 | ; CHECK-NEXT: --> {(%a + %b),+,3}<%loop2> |
| 157 | ; CHECK: %sum4 = add i32 %sum3, %phi6 |
| 158 | ; CHECK-NEXT: --> {(43 + %a + %b),+,6}<%loop2> |
| 159 | ; CHECK: %is2 = add i32 %sum4, %v2 |
| 160 | ; CHECK-NEXT: --> ({(43 + %a + %b),+,6}<%loop2> + %v2) |
| 161 | ; CHECK: %is3 = add i32 %v1, %sum2 |
| 162 | ; CHECK-NEXT: --> ({(6 + %a + %b),+,6}<%loop1> + %v1) |
| 163 | ; CHECK: %ec2 = add i32 %is1, %is3 |
| 164 | ; CHECK-NEXT: --> (2 * ({(6 + %a + %b),+,6}<%loop1> + %v1)) |
| 165 | ; CHECK: %s1 = add i32 %phi1, %is1 |
| 166 | ; CHECK-NEXT: --> ({(6 + (2 * %a) + %b),+,7}<%loop1> + %v1) |
| 167 | ; CHECK: %s2 = add i32 %is2, %phi4 |
| 168 | ; CHECK-NEXT: --> ({(43 + (2 * %a) + %b),+,7}<%loop2> + %v2) |
| 169 | ; CHECK: %s3 = add i32 %is1, %phi5 |
| 170 | ; CHECK-NEXT: --> {({(6 + (2 * %b) + %a),+,6}<%loop1> + %v1),+,2}<%loop2> |
| 171 | ; CHECK: %s4 = add i32 %phi2, %is2 |
| 172 | ; CHECK-NEXT: --> ({{{{}}(43 + (2 * %b) + %a),+,2}<%loop1>,+,6}<%loop2> + %v2) |
| 173 | ; CHECK: %s5 = add i32 %is1, %is2 |
| 174 | ; CHECK-NEXT: --> ({({(49 + (2 * %a) + (2 * %b)),+,6}<%loop1> + %v1),+,6}<%loop2> + %v2) |
| 175 | ; CHECK: %s6 = add i32 %is2, %is1 |
| 176 | ; CHECK-NEXT: --> ({({(49 + (2 * %a) + (2 * %b)),+,6}<%loop1> + %v1),+,6}<%loop2> + %v2) |
| 177 | |
| 178 | entry: |
| 179 | br label %loop1 |
| 180 | |
| 181 | loop1: |
| 182 | %phi1 = phi i32 [ %a, %entry ], [ %phi1.inc, %loop1 ] |
| 183 | %phi2 = phi i32 [ %b, %entry ], [ %phi2.inc, %loop1 ] |
| 184 | %phi3 = phi i32 [ 6, %entry ], [ %phi3.inc, %loop1 ] |
| 185 | %phi1.inc = add i32 %phi1, 1 |
| 186 | %phi2.inc = add i32 %phi2, 2 |
| 187 | %phi3.inc = add i32 %phi3, 3 |
| 188 | %v1 = load i32, i32* %p |
| 189 | %sum1 = add i32 %phi1, %phi2 |
| 190 | %sum2 = add i32 %sum1, %phi3 |
| 191 | %is1 = add i32 %sum2, %v1 |
| 192 | %cond1 = icmp ult i32 %is1, 1000 |
| 193 | br i1 %cond1, label %loop1, label %loop2 |
| 194 | |
| 195 | loop2: |
| 196 | %phi4 = phi i32 [ %a, %loop1 ], [ %phi4.inc, %loop2 ] |
| 197 | %phi5 = phi i32 [ %b, %loop1 ], [ %phi5.inc, %loop2 ] |
| 198 | %phi6 = phi i32 [ 43, %loop1 ], [ %phi6.inc, %loop2 ] |
| 199 | %phi4.inc = add i32 %phi4, 1 |
| 200 | %phi5.inc = add i32 %phi5, 2 |
| 201 | %phi6.inc = add i32 %phi6, 3 |
| 202 | %v2 = load i32, i32* %p |
| 203 | %sum3 = add i32 %phi4, %phi5 |
| 204 | %sum4 = add i32 %sum3, %phi6 |
| 205 | %is2 = add i32 %sum4, %v2 |
| 206 | %is3 = add i32 %v1, %sum2 |
| 207 | %ec2 = add i32 %is1, %is3 |
| 208 | %cond2 = icmp ult i32 %ec2, 1000 |
| 209 | br i1 %cond2, label %loop2, label %exit |
| 210 | |
| 211 | exit: |
| 212 | %s1 = add i32 %phi1, %is1 |
| 213 | %s2 = add i32 %is2, %phi4 |
| 214 | %s3 = add i32 %is1, %phi5 |
| 215 | %s4 = add i32 %phi2, %is2 |
| 216 | %s5 = add i32 %is1, %is2 |
| 217 | %s6 = add i32 %is2, %is1 |
| 218 | ret void |
| 219 | } |
| 220 | |
| 221 | ; Mix of previous use cases that demonstrates %s3 can be incorrectly treated as |
| 222 | ; a recurrence of loop1 because of operands order if we pick recurrencies in an |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 223 | ; incorrect order. It also shows that we cannot safely fold v1 (SCEVUnknown) |
| 224 | ; because we cannot prove for sure that it doesn't use Phis of loop 2. |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 225 | |
| 226 | define void @test_03(i32 %a, i32 %b, i32 %c, i32* %p) { |
| 227 | |
| 228 | ; CHECK-LABEL: Classifying expressions for: @test_03 |
| 229 | ; CHECK: %v1 = load i32, i32* %p |
| 230 | ; CHECK-NEXT: --> %v1 |
| 231 | ; CHECK: %s1 = add i32 %phi1, %v1 |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 232 | ; CHECK-NEXT: --> ({%a,+,1}<%loop1> + %v1) |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 233 | ; CHECK: %s2 = add i32 %s1, %b |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 234 | ; CHECK-NEXT: --> ({(%a + %b),+,1}<%loop1> + %v1) |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 235 | ; CHECK: %s3 = add i32 %s2, %phi2 |
| 236 | ; CHECK-NEXT: --> ({{{{}}((2 * %a) + %b),+,1}<%loop1>,+,2}<%loop2> + %v1) |
| 237 | |
| 238 | entry: |
| 239 | br label %loop1 |
| 240 | |
| 241 | loop1: |
| 242 | %phi1 = phi i32 [ %a, %entry ], [ %phi1.inc, %loop1 ] |
| 243 | %phi1.inc = add i32 %phi1, 1 |
| 244 | %cond1 = icmp ult i32 %phi1, %c |
| 245 | br i1 %cond1, label %loop1, label %loop2 |
| 246 | |
| 247 | loop2: |
| 248 | %phi2 = phi i32 [ %a, %loop1 ], [ %phi2.inc, %loop2 ] |
| 249 | %phi2.inc = add i32 %phi2, 2 |
| 250 | %v1 = load i32, i32* %p |
| 251 | %s1 = add i32 %phi1, %v1 |
| 252 | %s2 = add i32 %s1, %b |
| 253 | %s3 = add i32 %s2, %phi2 |
| 254 | %cond2 = icmp ult i32 %s3, %c |
| 255 | br i1 %cond2, label %loop2, label %exit |
| 256 | |
| 257 | exit: |
| 258 | |
| 259 | ret void |
| 260 | } |
| 261 | |
| 262 | ; Another mix of previous use cases that demonstrates that incorrect picking of |
| 263 | ; a loop for a recurrence may cause a crash of SCEV analysis. |
| 264 | define void @test_04() { |
| 265 | |
| 266 | ; CHECK-LABEL: Classifying expressions for: @test_04 |
| 267 | ; CHECK: %tmp = phi i64 [ 2, %bb ], [ %tmp4, %bb3 ] |
| 268 | ; CHECK-NEXT: --> {2,+,1}<nuw><nsw><%loop1> |
| 269 | ; CHECK: %tmp2 = trunc i64 %tmp to i32 |
| 270 | ; CHECK-NEXT: --> {2,+,1}<%loop1> |
| 271 | ; CHECK: %tmp4 = add nuw nsw i64 %tmp, 1 |
| 272 | ; CHECK-NEXT: --> {3,+,1}<nuw><%loop1> |
| 273 | ; CHECK: %tmp7 = phi i64 [ %tmp15, %loop2 ], [ 2, %loop1 ] |
| 274 | ; CHECK-NEXT: --> {2,+,1}<nuw><nsw><%loop2> |
| 275 | ; CHECK: %tmp10 = sub i64 %tmp9, %tmp7 |
| 276 | ; CHECK-NEXT: --> ((sext i8 %tmp8 to i64) + {-2,+,-1}<nw><%loop2>) |
| 277 | ; CHECK: %tmp11 = add i64 %tmp10, undef |
| 278 | ; CHECK-NEXT: --> ((sext i8 %tmp8 to i64) + {(-2 + undef),+,-1}<nw><%loop2>) |
| 279 | ; CHECK: %tmp13 = trunc i64 %tmp11 to i32 |
Justin Lebar | b326904 | 2018-06-14 17:13:35 +0000 | [diff] [blame] | 280 | ; CHECK-NEXT: --> ((sext i8 %tmp8 to i32) + {(-2 + (trunc i64 undef to i32)),+,-1}<%loop2>) |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 281 | ; CHECK: %tmp14 = sub i32 %tmp13, %tmp2 |
Justin Lebar | b326904 | 2018-06-14 17:13:35 +0000 | [diff] [blame] | 282 | ; `{{[{][{]}}` is the ugliness needed to match `{{` |
| 283 | ; CHECK-NEXT: --> ((sext i8 %tmp8 to i32) + {{[{][{]}}(-4 + (trunc i64 undef to i32)),+,-1}<%loop1>,+,-1}<%loop2>) |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 284 | ; CHECK: %tmp15 = add nuw nsw i64 %tmp7, 1 |
| 285 | ; CHECK-NEXT: --> {3,+,1}<nuw><nsw><%loop2> |
| 286 | |
| 287 | bb: |
| 288 | br label %loop1 |
| 289 | |
| 290 | loop1: |
| 291 | %tmp = phi i64 [ 2, %bb ], [ %tmp4, %bb3 ] |
| 292 | %tmp2 = trunc i64 %tmp to i32 |
| 293 | br i1 undef, label %loop2, label %bb3 |
| 294 | |
| 295 | bb3: |
| 296 | %tmp4 = add nuw nsw i64 %tmp, 1 |
| 297 | br label %loop1 |
| 298 | |
| 299 | bb5: |
| 300 | ret void |
| 301 | |
| 302 | loop2: |
| 303 | %tmp7 = phi i64 [ %tmp15, %loop2 ], [ 2, %loop1 ] |
| 304 | %tmp8 = load i8, i8 addrspace(1)* undef, align 1 |
| 305 | %tmp9 = sext i8 %tmp8 to i64 |
| 306 | %tmp10 = sub i64 %tmp9, %tmp7 |
| 307 | %tmp11 = add i64 %tmp10, undef |
| 308 | %tmp13 = trunc i64 %tmp11 to i32 |
| 309 | %tmp14 = sub i32 %tmp13, %tmp2 |
| 310 | %tmp15 = add nuw nsw i64 %tmp7, 1 |
| 311 | %tmp16 = icmp slt i64 %tmp15, %tmp |
| 312 | br i1 %tmp16, label %loop2, label %bb5 |
| 313 | } |
| 314 | |
| 315 | @A = weak global [1000 x i32] zeroinitializer, align 32 |
| 316 | |
| 317 | ; Demonstrate a situation when we can add two recs with different degrees from |
| 318 | ; the same loop. |
| 319 | define void @test_05(i32 %N) { |
| 320 | |
| 321 | ; CHECK-LABEL: Classifying expressions for: @test_05 |
| 322 | ; CHECK: %SQ = mul i32 %i.0, %i.0 |
| 323 | ; CHECK-NEXT: --> {4,+,5,+,2}<%bb3> |
| 324 | ; CHECK: %tmp4 = mul i32 %i.0, 2 |
Tim Shen | a064622 | 2018-07-13 23:58:46 +0000 | [diff] [blame] | 325 | ; CHECK-NEXT: --> {4,+,2}<nuw><nsw><%bb3> |
Max Kazantsev | b09b5db | 2017-05-16 07:27:06 +0000 | [diff] [blame] | 326 | ; CHECK: %tmp5 = sub i32 %SQ, %tmp4 |
| 327 | ; CHECK-NEXT: --> {0,+,3,+,2}<%bb3> |
| 328 | |
| 329 | entry: |
| 330 | %"alloca point" = bitcast i32 0 to i32 ; <i32> [#uses=0] |
| 331 | br label %bb3 |
| 332 | |
| 333 | bb: ; preds = %bb3 |
| 334 | %tmp = getelementptr [1000 x i32], [1000 x i32]* @A, i32 0, i32 %i.0 ; <i32*> [#uses=1] |
| 335 | store i32 123, i32* %tmp |
| 336 | %tmp2 = add i32 %i.0, 1 ; <i32> [#uses=1] |
| 337 | br label %bb3 |
| 338 | |
| 339 | bb3: ; preds = %bb, %entry |
| 340 | %i.0 = phi i32 [ 2, %entry ], [ %tmp2, %bb ] ; <i32> [#uses=3] |
| 341 | %SQ = mul i32 %i.0, %i.0 |
| 342 | %tmp4 = mul i32 %i.0, 2 |
| 343 | %tmp5 = sub i32 %SQ, %tmp4 |
| 344 | %tmp3 = icmp sle i32 %tmp5, 9999 ; <i1> [#uses=1] |
| 345 | br i1 %tmp3, label %bb, label %bb5 |
| 346 | |
| 347 | bb5: ; preds = %bb3 |
| 348 | br label %return |
| 349 | |
| 350 | return: ; preds = %bb5 |
| 351 | ret void |
| 352 | } |
| 353 | |
| 354 | ; Check that we can add Phis from different loops with different nesting, nested |
| 355 | ; loop comes first. |
| 356 | define void @test_06() { |
| 357 | |
| 358 | ; CHECK-LABEL: Classifying expressions for: @test_06 |
| 359 | ; CHECK: %s1 = add i32 %phi1, %phi2 |
| 360 | ; CHECK-NEXT: --> {{{{}}30,+,1}<%loop1>,+,2}<%loop2> |
| 361 | ; CHECK: %s2 = add i32 %phi2, %phi1 |
| 362 | ; CHECK-NEXT: --> {{{{}}30,+,1}<%loop1>,+,2}<%loop2> |
| 363 | ; CHECK: %s3 = add i32 %phi1, %phi3 |
| 364 | ; CHECK-NEXT: --> {{{{}}40,+,1}<%loop1>,+,3}<%loop3> |
| 365 | ; CHECK: %s4 = add i32 %phi3, %phi1 |
| 366 | ; CHECK-NEXT: --> {{{{}}40,+,1}<%loop1>,+,3}<%loop3> |
| 367 | ; CHECK: %s5 = add i32 %phi2, %phi3 |
| 368 | ; CHECK-NEXT: --> {{{{}}50,+,2}<%loop2>,+,3}<%loop3> |
| 369 | ; CHECK: %s6 = add i32 %phi3, %phi2 |
| 370 | ; CHECK-NEXT: --> {{{{}}50,+,2}<%loop2>,+,3}<%loop3> |
| 371 | |
| 372 | entry: |
| 373 | br label %loop1 |
| 374 | |
| 375 | loop1: |
| 376 | %phi1 = phi i32 [ 10, %entry ], [ %phi1.inc, %loop1.exit ] |
| 377 | br label %loop2 |
| 378 | |
| 379 | loop2: |
| 380 | %phi2 = phi i32 [ 20, %loop1 ], [ %phi2.inc, %loop2 ] |
| 381 | %phi2.inc = add i32 %phi2, 2 |
| 382 | %cond2 = icmp ult i32 %phi2.inc, 1000 |
| 383 | br i1 %cond2, label %loop2, label %loop1.exit |
| 384 | |
| 385 | loop1.exit: |
| 386 | %phi1.inc = add i32 %phi1, 1 |
| 387 | %cond1 = icmp ult i32 %phi1.inc, 1000 |
| 388 | br i1 %cond1, label %loop1, label %loop3 |
| 389 | |
| 390 | loop3: |
| 391 | %phi3 = phi i32 [ 30, %loop1.exit ], [ %phi3.inc, %loop3 ] |
| 392 | %phi3.inc = add i32 %phi3, 3 |
| 393 | %cond3 = icmp ult i32 %phi3.inc, 1000 |
| 394 | br i1 %cond3, label %loop3, label %exit |
| 395 | |
| 396 | exit: |
| 397 | %s1 = add i32 %phi1, %phi2 |
| 398 | %s2 = add i32 %phi2, %phi1 |
| 399 | %s3 = add i32 %phi1, %phi3 |
| 400 | %s4 = add i32 %phi3, %phi1 |
| 401 | %s5 = add i32 %phi2, %phi3 |
| 402 | %s6 = add i32 %phi3, %phi2 |
| 403 | ret void |
| 404 | } |
| 405 | |
| 406 | ; Check that we can add Phis from different loops with different nesting, nested |
| 407 | ; loop comes second. |
| 408 | define void @test_07() { |
| 409 | |
| 410 | ; CHECK-LABEL: Classifying expressions for: @test_07 |
| 411 | ; CHECK: %s1 = add i32 %phi1, %phi2 |
| 412 | ; CHECK-NEXT: --> {{{{}}30,+,1}<%loop1>,+,2}<%loop2> |
| 413 | ; CHECK: %s2 = add i32 %phi2, %phi1 |
| 414 | ; CHECK-NEXT: --> {{{{}}30,+,1}<%loop1>,+,2}<%loop2> |
| 415 | ; CHECK: %s3 = add i32 %phi1, %phi3 |
| 416 | ; CHECK-NEXT: --> {{{{}}40,+,3}<%loop3>,+,1}<%loop1> |
| 417 | ; CHECK: %s4 = add i32 %phi3, %phi1 |
| 418 | ; CHECK-NEXT: --> {{{{}}40,+,3}<%loop3>,+,1}<%loop1> |
| 419 | ; CHECK: %s5 = add i32 %phi2, %phi3 |
| 420 | ; CHECK-NEXT: --> {{{{}}50,+,3}<%loop3>,+,2}<%loop2> |
| 421 | ; CHECK: %s6 = add i32 %phi3, %phi2 |
| 422 | ; CHECK-NEXT: --> {{{{}}50,+,3}<%loop3>,+,2}<%loop2> |
| 423 | |
| 424 | entry: |
| 425 | br label %loop3 |
| 426 | |
| 427 | loop3: |
| 428 | %phi3 = phi i32 [ 30, %entry ], [ %phi3.inc, %loop3 ] |
| 429 | %phi3.inc = add i32 %phi3, 3 |
| 430 | %cond3 = icmp ult i32 %phi3.inc, 1000 |
| 431 | br i1 %cond3, label %loop3, label %loop1 |
| 432 | |
| 433 | loop1: |
| 434 | %phi1 = phi i32 [ 10, %loop3 ], [ %phi1.inc, %loop1.exit ] |
| 435 | br label %loop2 |
| 436 | |
| 437 | loop2: |
| 438 | %phi2 = phi i32 [ 20, %loop1 ], [ %phi2.inc, %loop2 ] |
| 439 | %phi2.inc = add i32 %phi2, 2 |
| 440 | %cond2 = icmp ult i32 %phi2.inc, 1000 |
| 441 | br i1 %cond2, label %loop2, label %loop1.exit |
| 442 | |
| 443 | loop1.exit: |
| 444 | %phi1.inc = add i32 %phi1, 1 |
| 445 | %cond1 = icmp ult i32 %phi1.inc, 1000 |
| 446 | br i1 %cond1, label %exit, label %loop1 |
| 447 | |
| 448 | exit: |
| 449 | %s1 = add i32 %phi1, %phi2 |
| 450 | %s2 = add i32 %phi2, %phi1 |
| 451 | %s3 = add i32 %phi1, %phi3 |
| 452 | %s4 = add i32 %phi3, %phi1 |
| 453 | %s5 = add i32 %phi2, %phi3 |
| 454 | %s6 = add i32 %phi3, %phi2 |
| 455 | ret void |
| 456 | } |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 457 | |
| 458 | ; Make sure that a complicated Phi does not get folded with rec's start value |
| 459 | ; of a loop which is above. |
| 460 | define void @test_08() { |
| 461 | |
| 462 | ; CHECK-LABEL: Classifying expressions for: @test_08 |
| 463 | ; CHECK: %tmp11 = add i64 %iv.2.2, %iv.2.1 |
| 464 | ; CHECK-NEXT: --> ({0,+,-1}<nsw><%loop_2> + %iv.2.1) |
| 465 | ; CHECK: %tmp12 = trunc i64 %tmp11 to i32 |
Justin Lebar | b326904 | 2018-06-14 17:13:35 +0000 | [diff] [blame] | 466 | ; CHECK-NEXT: --> ((trunc i64 %iv.2.1 to i32) + {0,+,-1}<%loop_2>) |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 467 | ; CHECK: %tmp14 = mul i32 %tmp12, %tmp7 |
Justin Lebar | b326904 | 2018-06-14 17:13:35 +0000 | [diff] [blame] | 468 | ; CHECK-NEXT: --> (((trunc i64 %iv.2.1 to i32) + {0,+,-1}<%loop_2>) * {-1,+,-1}<%loop_1>) |
Max Kazantsev | 4145032 | 2017-05-26 06:47:04 +0000 | [diff] [blame] | 469 | ; CHECK: %tmp16 = mul i64 %iv.2.1, %iv.1.1 |
| 470 | ; CHECK-NEXT: --> ({2,+,1}<nuw><nsw><%loop_1> * %iv.2.1) |
| 471 | |
| 472 | entry: |
| 473 | br label %loop_1 |
| 474 | |
| 475 | loop_1: |
| 476 | %iv.1.1 = phi i64 [ 2, %entry ], [ %iv.1.1.next, %loop_1_back_branch ] |
| 477 | %iv.1.2 = phi i32 [ -1, %entry ], [ %iv.1.2.next, %loop_1_back_branch ] |
| 478 | br label %loop_1_exit |
| 479 | |
| 480 | dead: |
| 481 | br label %loop_1_exit |
| 482 | |
| 483 | loop_1_exit: |
| 484 | %tmp5 = icmp sgt i64 %iv.1.1, 2 |
| 485 | br i1 %tmp5, label %loop_2_preheader, label %loop_1_back_branch |
| 486 | |
| 487 | loop_1_back_branch: |
| 488 | %iv.1.1.next = add nuw nsw i64 %iv.1.1, 1 |
| 489 | %iv.1.2.next = add nsw i32 %iv.1.2, 1 |
| 490 | br label %loop_1 |
| 491 | |
| 492 | loop_2_preheader: |
| 493 | %tmp6 = sub i64 1, %iv.1.1 |
| 494 | %tmp7 = trunc i64 %tmp6 to i32 |
| 495 | br label %loop_2 |
| 496 | |
| 497 | loop_2: |
| 498 | %iv.2.1 = phi i64 [ 0, %loop_2_preheader ], [ %tmp16, %loop_2 ] |
| 499 | %iv.2.2 = phi i64 [ 0, %loop_2_preheader ], [ %iv.2.2.next, %loop_2 ] |
| 500 | %iv.2.3 = phi i64 [ 2, %loop_2_preheader ], [ %iv.2.3.next, %loop_2 ] |
| 501 | %tmp11 = add i64 %iv.2.2, %iv.2.1 |
| 502 | %tmp12 = trunc i64 %tmp11 to i32 |
| 503 | %tmp14 = mul i32 %tmp12, %tmp7 |
| 504 | %tmp16 = mul i64 %iv.2.1, %iv.1.1 |
| 505 | %iv.2.3.next = add nuw nsw i64 %iv.2.3, 1 |
| 506 | %iv.2.2.next = add nsw i64 %iv.2.2, -1 |
| 507 | %tmp17 = icmp slt i64 %iv.2.3.next, %iv.1.1 |
| 508 | br i1 %tmp17, label %loop_2, label %exit |
| 509 | |
| 510 | exit: |
| 511 | %tmp10 = add i32 %iv.1.2, 3 |
| 512 | ret void |
| 513 | } |
Max Kazantsev | 23044fa | 2017-11-22 06:21:39 +0000 | [diff] [blame] | 514 | |
| 515 | define i64 @test_09(i32 %param) { |
| 516 | |
| 517 | ; CHECK-LABEL: Classifying expressions for: @test_09 |
| 518 | ; CHECK: %iv1 = phi i64 [ %iv1.next, %guarded ], [ 0, %outer.loop ] |
| 519 | ; CHECK-NEXT: --> {0,+,1}<nuw><nsw><%loop1> |
| 520 | ; CHECK: %iv1.trunc = trunc i64 %iv1 to i32 |
| 521 | ; CHECK-NEXT: --> {0,+,1}<%loop1> |
| 522 | ; CHECK: %iv1.next = add nuw nsw i64 %iv1, 1 |
| 523 | ; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%loop1> |
| 524 | ; CHECK: %iv2 = phi i32 [ %iv2.next, %loop2 ], [ %param, %loop2.preheader ] |
| 525 | ; CHECK-NEXT: --> {%param,+,1}<%loop2> |
| 526 | ; CHECK: %iv2.next = add i32 %iv2, 1 |
| 527 | ; CHECK-NEXT: --> {(1 + %param),+,1}<%loop2> |
| 528 | ; CHECK: %iv2.ext = sext i32 %iv2.next to i64 |
| 529 | ; CHECK-NEXT: --> (sext i32 {(1 + %param),+,1}<%loop2> to i64) |
| 530 | ; CHECK: %ret = mul i64 %iv1, %iv2.ext |
| 531 | ; CHECK-NEXT: --> ((sext i32 {(1 + %param),+,1}<%loop2> to i64) * {0,+,1}<nuw><nsw><%loop1>) |
| 532 | |
| 533 | entry: |
| 534 | br label %outer.loop |
| 535 | |
| 536 | outer.loop: ; preds = %loop2.exit, %entry |
| 537 | br label %loop1 |
| 538 | |
| 539 | loop1: ; preds = %guarded, %outer.loop |
| 540 | %iv1 = phi i64 [ %iv1.next, %guarded ], [ 0, %outer.loop ] |
| 541 | %iv1.trunc = trunc i64 %iv1 to i32 |
| 542 | %cond1 = icmp ult i64 %iv1, 100 |
| 543 | br i1 %cond1, label %guarded, label %deopt |
| 544 | |
| 545 | guarded: ; preds = %loop1 |
| 546 | %iv1.next = add nuw nsw i64 %iv1, 1 |
| 547 | %tmp16 = icmp slt i32 %iv1.trunc, 2 |
| 548 | br i1 %tmp16, label %loop1, label %loop2.preheader |
| 549 | |
| 550 | deopt: ; preds = %loop1 |
| 551 | unreachable |
| 552 | |
| 553 | loop2.preheader: ; preds = %guarded |
| 554 | br label %loop2 |
| 555 | |
| 556 | loop2: ; preds = %loop2, %loop2.preheader |
| 557 | %iv2 = phi i32 [ %iv2.next, %loop2 ], [ %param, %loop2.preheader ] |
| 558 | %iv2.next = add i32 %iv2, 1 |
| 559 | %cond2 = icmp slt i32 %iv2, %iv1.trunc |
| 560 | br i1 %cond2, label %loop2, label %exit |
| 561 | |
| 562 | exit: ; preds = %loop2.exit |
| 563 | %iv2.ext = sext i32 %iv2.next to i64 |
| 564 | %ret = mul i64 %iv1, %iv2.ext |
| 565 | ret i64 %ret |
| 566 | } |
| 567 | |
| 568 | define i64 @test_10(i32 %param) { |
| 569 | |
| 570 | ; CHECK-LABEL: Classifying expressions for: @test_10 |
| 571 | ; CHECK: %uncle = phi i64 [ %uncle.outer.next, %uncle.loop.backedge ], [ 0, %outer.loop ] |
| 572 | ; CHECK-NEXT: --> {0,+,1}<%uncle.loop> |
| 573 | ; CHECK: %iv1 = phi i64 [ %iv1.next, %guarded ], [ 0, %uncle.loop ] |
| 574 | ; CHECK-NEXT: --> {0,+,1}<nuw><nsw><%loop1> |
| 575 | ; CHECK: %iv1.trunc = trunc i64 %iv1 to i32 |
| 576 | ; CHECK-NEXT: --> {0,+,1}<%loop1> |
| 577 | ; CHECK: %iv1.next = add nuw nsw i64 %iv1, 1 |
| 578 | ; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%loop1> |
| 579 | ; CHECK: %uncle.outer.next = add i64 %uncle, 1 |
| 580 | ; CHECK-NEXT: --> {1,+,1}<%uncle.loop> |
| 581 | ; CHECK: %iv2 = phi i32 [ %iv2.next, %loop2 ], [ %param, %loop2.preheader ] |
| 582 | ; CHECK-NEXT: --> {%param,+,1}<%loop2> |
| 583 | ; CHECK: %iv2.next = add i32 %iv2, 1 |
| 584 | ; CHECK-NEXT: --> {(1 + %param),+,1}<%loop2> |
| 585 | ; CHECK: %iv2.ext = sext i32 %iv2.next to i64 |
| 586 | ; CHECK-NEXT: --> (sext i32 {(1 + %param),+,1}<%loop2> to i64) |
| 587 | ; CHECK: %ret = mul i64 %iv1, %iv2.ext |
| 588 | ; CHECK-NEXT: --> ((sext i32 {(1 + %param),+,1}<%loop2> to i64) * {0,+,1}<nuw><nsw><%loop1>) |
| 589 | |
| 590 | entry: |
| 591 | br label %outer.loop |
| 592 | |
| 593 | outer.loop: ; preds = %entry |
| 594 | br label %uncle.loop |
| 595 | |
| 596 | uncle.loop: ; preds = %uncle.loop.backedge, %outer.loop |
| 597 | %uncle = phi i64 [ %uncle.outer.next, %uncle.loop.backedge ], [ 0, %outer.loop ] |
| 598 | br label %loop1 |
| 599 | |
| 600 | loop1: ; preds = %guarded, %uncle.loop |
| 601 | %iv1 = phi i64 [ %iv1.next, %guarded ], [ 0, %uncle.loop ] |
| 602 | %iv1.trunc = trunc i64 %iv1 to i32 |
| 603 | %cond1 = icmp ult i64 %iv1, 100 |
| 604 | br i1 %cond1, label %guarded, label %deopt |
| 605 | |
| 606 | guarded: ; preds = %loop1 |
| 607 | %iv1.next = add nuw nsw i64 %iv1, 1 |
| 608 | %tmp16 = icmp slt i32 %iv1.trunc, 2 |
| 609 | br i1 %tmp16, label %loop1, label %uncle.loop.backedge |
| 610 | |
| 611 | uncle.loop.backedge: ; preds = %guarded |
| 612 | %uncle.outer.next = add i64 %uncle, 1 |
| 613 | %cond.uncle = icmp ult i64 %uncle, 120 |
| 614 | br i1 %cond.uncle, label %loop2.preheader, label %uncle.loop |
| 615 | |
| 616 | deopt: ; preds = %loop1 |
| 617 | unreachable |
| 618 | |
| 619 | loop2.preheader: ; preds = %uncle.loop.backedge |
| 620 | br label %loop2 |
| 621 | |
| 622 | loop2: ; preds = %loop2, %loop2.preheader |
| 623 | %iv2 = phi i32 [ %iv2.next, %loop2 ], [ %param, %loop2.preheader ] |
| 624 | %iv2.next = add i32 %iv2, 1 |
| 625 | %cond2 = icmp slt i32 %iv2, %iv1.trunc |
| 626 | br i1 %cond2, label %loop2, label %exit |
| 627 | |
| 628 | exit: ; preds = %loop2 |
| 629 | %iv2.ext = sext i32 %iv2.next to i64 |
| 630 | %ret = mul i64 %iv1, %iv2.ext |
| 631 | ret i64 %ret |
| 632 | } |