1; RUN: opt < %s -indvars -S | FileCheck %s 2; 3; Make sure that indvars isn't inserting canonical IVs. 4; This is kinda hard to do until linear function test replacement is removed. 5 6target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64" 7 8define i32 @sum(i32* %arr, i32 %n) nounwind { 9entry: 10 %precond = icmp slt i32 0, %n 11 br i1 %precond, label %ph, label %return 12 13ph: 14 br label %loop 15 16; CHECK: loop: 17; 18; We should only have 2 IVs. 19; CHECK: phi 20; CHECK: phi 21; CHECK-NOT: phi 22; 23; sext should be eliminated while preserving gep inboundsness. 24; CHECK-NOT: sext 25; CHECK: getelementptr inbounds 26; CHECK: exit: 27loop: 28 %i.02 = phi i32 [ 0, %ph ], [ %iinc, %loop ] 29 %s.01 = phi i32 [ 0, %ph ], [ %sinc, %loop ] 30 %ofs = sext i32 %i.02 to i64 31 %adr = getelementptr inbounds i32* %arr, i64 %ofs 32 %val = load i32* %adr 33 %sinc = add nsw i32 %s.01, %val 34 %iinc = add nsw i32 %i.02, 1 35 %cond = icmp slt i32 %iinc, %n 36 br i1 %cond, label %loop, label %exit 37 38exit: 39 %s.lcssa = phi i32 [ %sinc, %loop ] 40 br label %return 41 42return: 43 %s.0.lcssa = phi i32 [ %s.lcssa, %exit ], [ 0, %entry ] 44 ret i32 %s.0.lcssa 45} 46 47define i64 @suml(i32* %arr, i32 %n) nounwind { 48entry: 49 %precond = icmp slt i32 0, %n 50 br i1 %precond, label %ph, label %return 51 52ph: 53 br label %loop 54 55; CHECK: loop: 56; 57; We should only have 2 IVs. 58; CHECK: phi 59; CHECK: phi 60; CHECK-NOT: phi 61; 62; %ofs sext should be eliminated while preserving gep inboundsness. 63; CHECK-NOT: sext 64; CHECK: getelementptr inbounds 65; %vall sext should obviously not be eliminated 66; CHECK: sext 67; CHECK: exit: 68loop: 69 %i.02 = phi i32 [ 0, %ph ], [ %iinc, %loop ] 70 %s.01 = phi i64 [ 0, %ph ], [ %sinc, %loop ] 71 %ofs = sext i32 %i.02 to i64 72 %adr = getelementptr inbounds i32* %arr, i64 %ofs 73 %val = load i32* %adr 74 %vall = sext i32 %val to i64 75 %sinc = add nsw i64 %s.01, %vall 76 %iinc = add nsw i32 %i.02, 1 77 %cond = icmp slt i32 %iinc, %n 78 br i1 %cond, label %loop, label %exit 79 80exit: 81 %s.lcssa = phi i64 [ %sinc, %loop ] 82 br label %return 83 84return: 85 %s.0.lcssa = phi i64 [ %s.lcssa, %exit ], [ 0, %entry ] 86 ret i64 %s.0.lcssa 87} 88 89define void @outofbounds(i32* %first, i32* %last, i32 %idx) nounwind { 90 %precond = icmp ne i32* %first, %last 91 br i1 %precond, label %ph, label %return 92 93; CHECK: ph: 94; It's not indvars' job to perform LICM on %ofs 95; CHECK-NOT: sext 96ph: 97 br label %loop 98 99; CHECK: loop: 100; 101; Preserve exactly one pointer type IV. 102; CHECK: phi i32* 103; CHECK-NOT: phi 104; 105; Don't create any extra adds. 106; CHECK-NOT: add 107; 108; Preserve gep inboundsness, and don't factor it. 109; CHECK: getelementptr inbounds i32* %ptriv, i32 1 110; CHECK-NOT: add 111; CHECK: exit: 112loop: 113 %ptriv = phi i32* [ %first, %ph ], [ %ptrpost, %loop ] 114 %ofs = sext i32 %idx to i64 115 %adr = getelementptr inbounds i32* %ptriv, i64 %ofs 116 store i32 3, i32* %adr 117 %ptrpost = getelementptr inbounds i32* %ptriv, i32 1 118 %cond = icmp ne i32* %ptrpost, %last 119 br i1 %cond, label %loop, label %exit 120 121exit: 122 br label %return 123 124return: 125 ret void 126} 127 128%structI = type { i32 } 129 130define void @bitcastiv(i32 %start, i32 %limit, i32 %step, %structI* %base) 131nounwind 132{ 133entry: 134 br label %loop 135 136; CHECK: loop: 137; 138; Preserve casts 139; CHECK: phi i32 140; CHECK: bitcast 141; CHECK: getelementptr 142; CHECK: exit: 143loop: 144 %iv = phi i32 [%start, %entry], [%next, %loop] 145 %p = phi %structI* [%base, %entry], [%pinc, %loop] 146 %adr = getelementptr %structI* %p, i32 0, i32 0 147 store i32 3, i32* %adr 148 %pp = bitcast %structI* %p to i32* 149 store i32 4, i32* %pp 150 %pinc = getelementptr %structI* %p, i32 1 151 %next = add i32 %iv, 1 152 %cond = icmp ne i32 %next, %limit 153 br i1 %cond, label %loop, label %exit 154 155exit: 156 ret void 157} 158 159define void @maxvisitor(i32 %limit, i32* %base) nounwind { 160entry: 161 br label %loop 162 163; Test inserting a truncate at a phi use. 164; 165; CHECK: loop: 166; CHECK: phi i64 167; CHECK: trunc 168; CHECK: exit: 169loop: 170 %idx = phi i32 [ 0, %entry ], [ %idx.next, %loop.inc ] 171 %max = phi i32 [ 0, %entry ], [ %max.next, %loop.inc ] 172 %idxprom = sext i32 %idx to i64 173 %adr = getelementptr inbounds i32* %base, i64 %idxprom 174 %val = load i32* %adr 175 %cmp19 = icmp sgt i32 %val, %max 176 br i1 %cmp19, label %if.then, label %if.else 177 178if.then: 179 br label %loop.inc 180 181if.else: 182 br label %loop.inc 183 184loop.inc: 185 %max.next = phi i32 [ %idx, %if.then ], [ %max, %if.else ] 186 %idx.next = add nsw i32 %idx, 1 187 %cmp = icmp slt i32 %idx.next, %limit 188 br i1 %cmp, label %loop, label %exit 189 190exit: 191 ret void 192} 193 194define void @identityphi(i32 %limit) nounwind { 195entry: 196 br label %loop 197 198; Test an edge case of removing an identity phi that directly feeds 199; back to the loop iv. 200; 201; CHECK: loop: 202; CHECK-NOT: phi 203; CHECK: exit: 204loop: 205 %iv = phi i32 [ 0, %entry], [ %iv.next, %control ] 206 br i1 undef, label %if.then, label %control 207 208if.then: 209 br label %control 210 211control: 212 %iv.next = phi i32 [ %iv, %loop ], [ undef, %if.then ] 213 %cmp = icmp slt i32 %iv.next, %limit 214 br i1 %cmp, label %loop, label %exit 215 216exit: 217 ret void 218} 219 220define i64 @cloneOr(i32 %limit, i64* %base) nounwind { 221entry: 222 ; ensure that the loop can't overflow 223 %halfLim = ashr i32 %limit, 2 224 br label %loop 225 226; This test originally checked that the OR instruction was cloned. Now the 227; ScalarEvolution is able to understand the loop evolution and that '%iv' at the 228; end of the loop is an even value. Thus '%val' is computed at the end of the 229; loop and the OR instruction is replaced by an ADD keeping the result 230; equivalent. 231; 232; CHECK: loop: 233; CHECK: phi i64 234; CHECK-NOT: sext 235; CHECK: icmp slt i32 236; CHECK: exit: 237; CHECK: add i64 238loop: 239 %iv = phi i32 [ 0, %entry], [ %iv.next, %loop ] 240 %t1 = sext i32 %iv to i64 241 %adr = getelementptr i64* %base, i64 %t1 242 %val = load i64* %adr 243 %t2 = or i32 %iv, 1 244 %t3 = sext i32 %t2 to i64 245 %iv.next = add i32 %iv, 2 246 %cmp = icmp slt i32 %iv.next, %halfLim 247 br i1 %cmp, label %loop, label %exit 248 249exit: 250 %result = and i64 %val, %t3 251 ret i64 %result 252} 253 254; The i induction variable looks like a wrap-around, but it really is just 255; a simple affine IV. Make sure that indvars simplifies through. 256define i32 @indirectRecurrence() nounwind { 257entry: 258 br label %loop 259 260; ReplaceLoopExitValue should fold the return value to constant 9. 261; CHECK: loop: 262; CHECK: phi i32 263; CHECK: ret i32 9 264loop: 265 %j.0 = phi i32 [ 1, %entry ], [ %j.next, %cond_true ] 266 %i.0 = phi i32 [ 0, %entry ], [ %j.0, %cond_true ] 267 %tmp = icmp ne i32 %j.0, 10 268 br i1 %tmp, label %cond_true, label %return 269 270cond_true: 271 %j.next = add i32 %j.0, 1 272 br label %loop 273 274return: 275 ret i32 %i.0 276} 277 278; Eliminate the congruent phis j, k, and l. 279; Eliminate the redundant IV increments k.next and l.next. 280; Two phis should remain, one starting at %init, and one at %init1. 281; Two increments should remain, one by %step and one by %step1. 282; CHECK: loop: 283; CHECK: phi i32 284; CHECK: phi i32 285; CHECK-NOT: phi 286; CHECK: add i32 287; CHECK: add i32 288; CHECK: add i32 289; CHECK-NOT: add 290; CHECK: return: 291; 292; Five live-outs should remain. 293; CHECK: lcssa = phi 294; CHECK: lcssa = phi 295; CHECK: lcssa = phi 296; CHECK: lcssa = phi 297; CHECK: lcssa = phi 298; CHECK-NOT: phi 299; CHECK: ret 300define i32 @isomorphic(i32 %init, i32 %step, i32 %lim) nounwind { 301entry: 302 %step1 = add i32 %step, 1 303 %init1 = add i32 %init, %step1 304 %l.0 = sub i32 %init1, %step1 305 br label %loop 306 307loop: 308 %ii = phi i32 [ %init1, %entry ], [ %ii.next, %loop ] 309 %i = phi i32 [ %init, %entry ], [ %ii, %loop ] 310 %j = phi i32 [ %init, %entry ], [ %j.next, %loop ] 311 %k = phi i32 [ %init1, %entry ], [ %k.next, %loop ] 312 %l = phi i32 [ %l.0, %entry ], [ %l.next, %loop ] 313 %ii.next = add i32 %ii, %step1 314 %j.next = add i32 %j, %step1 315 %k.next = add i32 %k, %step1 316 %l.step = add i32 %l, %step 317 %l.next = add i32 %l.step, 1 318 %cmp = icmp ne i32 %ii.next, %lim 319 br i1 %cmp, label %loop, label %return 320 321return: 322 %sum1 = add i32 %i, %j.next 323 %sum2 = add i32 %sum1, %k.next 324 %sum3 = add i32 %sum1, %l.step 325 %sum4 = add i32 %sum1, %l.next 326 ret i32 %sum4 327} 328 329; Test a GEP IV that is derived from another GEP IV by a nop gep that 330; lowers the type without changing the expression. 331%structIF = type { i32, float } 332 333define void @congruentgepiv(%structIF* %base) nounwind uwtable ssp { 334entry: 335 %first = getelementptr inbounds %structIF* %base, i64 0, i32 0 336 br label %loop 337 338; CHECK: loop: 339; CHECK: phi %structIF* 340; CHECK-NOT: phi 341; CHECK: getelementptr inbounds 342; CHECK-NOT: getelementptr 343; CHECK: exit: 344loop: 345 %ptr.iv = phi %structIF* [ %ptr.inc, %latch ], [ %base, %entry ] 346 %next = phi i32* [ %next.inc, %latch ], [ %first, %entry ] 347 store i32 4, i32* %next 348 br i1 undef, label %latch, label %exit 349 350latch: ; preds = %for.inc50.i 351 %ptr.inc = getelementptr inbounds %structIF* %ptr.iv, i64 1 352 %next.inc = getelementptr inbounds %structIF* %ptr.inc, i64 0, i32 0 353 br label %loop 354 355exit: 356 ret void 357} 358 359; Test a widened IV that is used by a phi on different paths within the loop. 360; 361; CHECK: for.body: 362; CHECK: phi i64 363; CHECK: trunc i64 364; CHECK: if.then: 365; CHECK: for.inc: 366; CHECK: phi i32 367; CHECK: for.end: 368define void @phiUsesTrunc() nounwind { 369entry: 370 br i1 undef, label %for.body, label %for.end 371 372for.body: 373 %iv = phi i32 [ %inc, %for.inc ], [ 1, %entry ] 374 br i1 undef, label %if.then, label %if.else 375 376if.then: 377 br i1 undef, label %if.then33, label %for.inc 378 379if.then33: 380 br label %for.inc 381 382if.else: 383 br i1 undef, label %if.then97, label %for.inc 384 385if.then97: 386 %idxprom100 = sext i32 %iv to i64 387 br label %for.inc 388 389for.inc: 390 %kmin.1 = phi i32 [ %iv, %if.then33 ], [ 0, %if.then ], [ %iv, %if.then97 ], [ 0, %if.else ] 391 %inc = add nsw i32 %iv, 1 392 br i1 undef, label %for.body, label %for.end 393 394for.end: 395 ret void 396} 397