1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -indvars -S | FileCheck %s 3 4target datalayout = "e-m:e-i64:64-p:64:64:64-n8:16:32:64-S128" 5 6; When widening IV and its users, trunc and zext/sext are not needed 7; if the original 32-bit user is known to be non-negative, whether 8; the IV is considered signed or unsigned. 9define void @foo(i32* %A, i32* %B, i32* %C, i32 %N) { 10; CHECK-LABEL: @foo( 11; CHECK-NEXT: entry: 12; CHECK-NEXT: [[CMP1:%.*]] = icmp slt i32 0, [[N:%.*]] 13; CHECK-NEXT: br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]] 14; CHECK: for.body.lr.ph: 15; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64 16; CHECK-NEXT: br label [[FOR_BODY:%.*]] 17; CHECK: for.body: 18; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ 0, [[FOR_BODY_LR_PH]] ] 19; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDVARS_IV]] 20; CHECK-NEXT: [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4 21; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 2 22; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[C:%.*]], i64 [[TMP1]] 23; CHECK-NEXT: [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4 24; CHECK-NEXT: [[ADD3:%.*]] = add nsw i32 [[TMP0]], [[TMP2]] 25; CHECK-NEXT: [[TMP3:%.*]] = trunc i64 [[TMP1]] to i32 26; CHECK-NEXT: [[DIV0:%.*]] = udiv i32 5, [[TMP3]] 27; CHECK-NEXT: [[ADD4:%.*]] = add nsw i32 [[ADD3]], [[DIV0]] 28; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]] 29; CHECK-NEXT: store i32 [[ADD4]], i32* [[ARRAYIDX5]], align 4 30; CHECK-NEXT: br label [[FOR_INC]] 31; CHECK: for.inc: 32; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 33; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 34; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]] 35; CHECK: for.cond.for.end_crit_edge: 36; CHECK-NEXT: br label [[FOR_END]] 37; CHECK: for.end: 38; CHECK-NEXT: ret void 39; 40entry: 41 %cmp1 = icmp slt i32 0, %N 42 br i1 %cmp1, label %for.body.lr.ph, label %for.end 43 44for.body.lr.ph: ; preds = %entry 45 br label %for.body 46 47for.body: ; preds = %for.body.lr.ph, %for.inc 48 %i.02 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ] 49 %idxprom = sext i32 %i.02 to i64 50 %arrayidx = getelementptr inbounds i32, i32* %B, i64 %idxprom 51 %0 = load i32, i32* %arrayidx, align 4 52 %add = add nsw i32 %i.02, 2 53 %idxprom1 = zext i32 %add to i64 54 %arrayidx2 = getelementptr inbounds i32, i32* %C, i64 %idxprom1 55 %1 = load i32, i32* %arrayidx2, align 4 56 %add3 = add nsw i32 %0, %1 57 %div0 = udiv i32 5, %add 58 %add4 = add nsw i32 %add3, %div0 59 %idxprom4 = zext i32 %i.02 to i64 60 %arrayidx5 = getelementptr inbounds i32, i32* %A, i64 %idxprom4 61 store i32 %add4, i32* %arrayidx5, align 4 62 br label %for.inc 63 64for.inc: ; preds = %for.body 65 %inc = add nsw i32 %i.02, 1 66 %cmp = icmp slt i32 %inc, %N 67 br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge 68 69for.cond.for.end_crit_edge: ; preds = %for.inc 70 br label %for.end 71 72for.end: ; preds = %for.cond.for.end_crit_edge, %entry 73 ret void 74} 75 76define void @foo1(i32* %A, i32* %B, i32* %C, i32 %N) { 77; CHECK-LABEL: @foo1( 78; CHECK-NEXT: entry: 79; CHECK-NEXT: [[CMP1:%.*]] = icmp slt i32 0, [[N:%.*]] 80; CHECK-NEXT: br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]] 81; CHECK: for.body.lr.ph: 82; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64 83; CHECK-NEXT: br label [[FOR_BODY:%.*]] 84; CHECK: for.body: 85; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ 0, [[FOR_BODY_LR_PH]] ] 86; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDVARS_IV]] 87; CHECK-NEXT: [[TMP0:%.*]] = load i32, i32* [[ARRAYIDX]], align 4 88; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i64 [[INDVARS_IV]], 2 89; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, i32* [[C:%.*]], i64 [[TMP1]] 90; CHECK-NEXT: [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4 91; CHECK-NEXT: [[ADD3:%.*]] = add nsw i32 [[TMP0]], [[TMP2]] 92; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]] 93; CHECK-NEXT: store i32 [[ADD3]], i32* [[ARRAYIDX5]], align 4 94; CHECK-NEXT: br label [[FOR_INC]] 95; CHECK: for.inc: 96; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 97; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 98; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]] 99; CHECK: for.cond.for.end_crit_edge: 100; CHECK-NEXT: br label [[FOR_END]] 101; CHECK: for.end: 102; CHECK-NEXT: ret void 103; 104entry: 105 %cmp1 = icmp slt i32 0, %N 106 br i1 %cmp1, label %for.body.lr.ph, label %for.end 107 108for.body.lr.ph: ; preds = %entry 109 br label %for.body 110 111for.body: ; preds = %for.body.lr.ph, %for.inc 112 %i.02 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ] 113 %idxprom = zext i32 %i.02 to i64 114 %arrayidx = getelementptr inbounds i32, i32* %B, i64 %idxprom 115 %0 = load i32, i32* %arrayidx, align 4 116 %add = add nsw i32 %i.02, 2 117 %idxprom1 = sext i32 %add to i64 118 %arrayidx2 = getelementptr inbounds i32, i32* %C, i64 %idxprom1 119 %1 = load i32, i32* %arrayidx2, align 4 120 %add3 = add nsw i32 %0, %1 121 %idxprom4 = sext i32 %i.02 to i64 122 %arrayidx5 = getelementptr inbounds i32, i32* %A, i64 %idxprom4 123 store i32 %add3, i32* %arrayidx5, align 4 124 br label %for.inc 125 126for.inc: ; preds = %for.body 127 %inc = add nsw i32 %i.02, 1 128 %cmp = icmp slt i32 %inc, %N 129 br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge 130 131for.cond.for.end_crit_edge: ; preds = %for.inc 132 br label %for.end 133 134for.end: ; preds = %for.cond.for.end_crit_edge, %entry 135 ret void 136} 137 138 139@a = common global [100 x i32] zeroinitializer, align 16 140@b = common global [100 x i32] zeroinitializer, align 16 141 142define i32 @foo2(i32 %M) { 143; CHECK-LABEL: @foo2( 144; CHECK-NEXT: entry: 145; CHECK-NEXT: [[CMP1:%.*]] = icmp slt i32 0, [[M:%.*]] 146; CHECK-NEXT: br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]] 147; CHECK: for.body.lr.ph: 148; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[M]] to i64 149; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[M]] to i64 150; CHECK-NEXT: br label [[FOR_BODY:%.*]] 151; CHECK: for.body: 152; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ 0, [[FOR_BODY_LR_PH]] ] 153; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 [[INDVARS_IV]] 154; CHECK-NEXT: [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX]], align 4 155; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @b, i64 0, i64 [[INDVARS_IV]] 156; CHECK-NEXT: [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4 157; CHECK-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP1]], [[TMP2]] 158; CHECK-NEXT: [[TMP3:%.*]] = add nsw i64 [[INDVARS_IV]], [[TMP0]] 159; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 [[TMP3]] 160; CHECK-NEXT: store i32 [[ADD]], i32* [[ARRAYIDX5]], align 4 161; CHECK-NEXT: br label [[FOR_INC]] 162; CHECK: for.inc: 163; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 164; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 165; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]] 166; CHECK: for.cond.for.end_crit_edge: 167; CHECK-NEXT: br label [[FOR_END]] 168; CHECK: for.end: 169; CHECK-NEXT: [[CALL:%.*]] = call i32 @dummy(i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), i32* getelementptr inbounds ([100 x i32], [100 x i32]* @b, i32 0, i32 0)) 170; CHECK-NEXT: ret i32 0 171; 172entry: 173 %cmp1 = icmp slt i32 0, %M 174 br i1 %cmp1, label %for.body.lr.ph, label %for.end 175 176for.body.lr.ph: ; preds = %entry 177 br label %for.body 178 179for.body: ; preds = %for.body.lr.ph, %for.inc 180 %i.02 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ] 181 %idxprom = zext i32 %i.02 to i64 182 %arrayidx = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 %idxprom 183 %0 = load i32, i32* %arrayidx, align 4 184 %idxprom1 = sext i32 %i.02 to i64 185 %arrayidx2 = getelementptr inbounds [100 x i32], [100 x i32]* @b, i64 0, i64 %idxprom1 186 %1 = load i32, i32* %arrayidx2, align 4 187 %add = add nsw i32 %0, %1 188 %add3 = add nsw i32 %i.02, %M 189 %idxprom4 = sext i32 %add3 to i64 190 %arrayidx5 = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 %idxprom4 191 store i32 %add, i32* %arrayidx5, align 4 192 br label %for.inc 193 194for.inc: ; preds = %for.body 195 %inc = add nsw i32 %i.02, 1 196 %cmp = icmp slt i32 %inc, %M 197 br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge 198 199for.cond.for.end_crit_edge: ; preds = %for.inc 200 br label %for.end 201 202for.end: ; preds = %for.cond.for.end_crit_edge, %entry 203 %call = call i32 @dummy(i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), i32* getelementptr inbounds ([100 x i32], [100 x i32]* @b, i32 0, i32 0)) 204 ret i32 0 205} 206 207declare i32 @dummy(i32*, i32*) 208 209; A case where zext should not be eliminated when its operands could only be extended by sext. 210define i32 @foo3(i32 %M) { 211; CHECK-LABEL: @foo3( 212; CHECK-NEXT: entry: 213; CHECK-NEXT: [[CMP1:%.*]] = icmp slt i32 0, [[M:%.*]] 214; CHECK-NEXT: br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]] 215; CHECK: for.body.lr.ph: 216; CHECK-NEXT: [[TMP0:%.*]] = sext i32 [[M]] to i64 217; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[M]] to i64 218; CHECK-NEXT: br label [[FOR_BODY:%.*]] 219; CHECK: for.body: 220; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ 0, [[FOR_BODY_LR_PH]] ] 221; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 [[INDVARS_IV]] 222; CHECK-NEXT: [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX]], align 4 223; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @b, i64 0, i64 [[INDVARS_IV]] 224; CHECK-NEXT: [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX2]], align 4 225; CHECK-NEXT: [[ADD:%.*]] = add nsw i32 [[TMP1]], [[TMP2]] 226; CHECK-NEXT: [[TMP3:%.*]] = add nsw i64 [[INDVARS_IV]], [[TMP0]] 227; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32 228; CHECK-NEXT: [[IDXPROM4:%.*]] = zext i32 [[TMP4]] to i64 229; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 [[IDXPROM4]] 230; CHECK-NEXT: store i32 [[ADD]], i32* [[ARRAYIDX5]], align 4 231; CHECK-NEXT: br label [[FOR_INC]] 232; CHECK: for.inc: 233; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 234; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 235; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]] 236; CHECK: for.cond.for.end_crit_edge: 237; CHECK-NEXT: br label [[FOR_END]] 238; CHECK: for.end: 239; CHECK-NEXT: [[CALL:%.*]] = call i32 @dummy(i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), i32* getelementptr inbounds ([100 x i32], [100 x i32]* @b, i32 0, i32 0)) 240; CHECK-NEXT: ret i32 0 241; 242entry: 243 %cmp1 = icmp slt i32 0, %M 244 br i1 %cmp1, label %for.body.lr.ph, label %for.end 245 246for.body.lr.ph: ; preds = %entry 247 br label %for.body 248 249for.body: ; preds = %for.body.lr.ph, %for.inc 250 %i.02 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ] 251 %idxprom = sext i32 %i.02 to i64 252 %arrayidx = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 %idxprom 253 %0 = load i32, i32* %arrayidx, align 4 254 %idxprom1 = sext i32 %i.02 to i64 255 %arrayidx2 = getelementptr inbounds [100 x i32], [100 x i32]* @b, i64 0, i64 %idxprom1 256 %1 = load i32, i32* %arrayidx2, align 4 257 %add = add nsw i32 %0, %1 258 %add3 = add nsw i32 %i.02, %M 259 %idxprom4 = zext i32 %add3 to i64 260 %arrayidx5 = getelementptr inbounds [100 x i32], [100 x i32]* @a, i64 0, i64 %idxprom4 261 store i32 %add, i32* %arrayidx5, align 4 262 br label %for.inc 263 264for.inc: ; preds = %for.body 265 %inc = add nsw i32 %i.02, 1 266 %cmp = icmp slt i32 %inc, %M 267 br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge 268 269for.cond.for.end_crit_edge: ; preds = %for.inc 270 br label %for.end 271 272for.end: ; preds = %for.cond.for.end_crit_edge, %entry 273 %call = call i32 @dummy(i32* getelementptr inbounds ([100 x i32], [100 x i32]* @a, i32 0, i32 0), i32* getelementptr inbounds ([100 x i32], [100 x i32]* @b, i32 0, i32 0)) 274 ret i32 0 275} 276 277%struct.image = type {i32, i32} 278define i32 @foo4(%struct.image* %input, i32 %length, i32* %in) { 279; CHECK-LABEL: @foo4( 280; CHECK-NEXT: entry: 281; CHECK-NEXT: [[STRIDE:%.*]] = getelementptr inbounds [[STRUCT_IMAGE:%.*]], %struct.image* [[INPUT:%.*]], i64 0, i32 1 282; CHECK-NEXT: [[TMP0:%.*]] = load i32, i32* [[STRIDE]], align 4 283; CHECK-NEXT: [[CMP17:%.*]] = icmp sgt i32 [[LENGTH:%.*]], 1 284; CHECK-NEXT: br i1 [[CMP17]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]] 285; CHECK: for.body.lr.ph: 286; CHECK-NEXT: [[CHANNEL:%.*]] = getelementptr inbounds [[STRUCT_IMAGE]], %struct.image* [[INPUT]], i64 0, i32 0 287; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[TMP0]] to i64 288; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[LENGTH]] to i64 289; CHECK-NEXT: br label [[FOR_BODY:%.*]] 290; CHECK: for.cond.cleanup.loopexit: 291; CHECK-NEXT: [[TMP2:%.*]] = phi i32 [ [[TMP10:%.*]], [[FOR_BODY]] ] 292; CHECK-NEXT: br label [[FOR_COND_CLEANUP]] 293; CHECK: for.cond.cleanup: 294; CHECK-NEXT: [[TMP3:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[TMP2]], [[FOR_COND_CLEANUP_LOOPEXIT:%.*]] ] 295; CHECK-NEXT: ret i32 [[TMP3]] 296; CHECK: for.body: 297; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ 1, [[FOR_BODY_LR_PH]] ] 298; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 299; CHECK-NEXT: [[TMP4:%.*]] = load i32, i32* [[CHANNEL]], align 8 300; CHECK-NEXT: [[TMP5:%.*]] = sext i32 [[TMP4]] to i64 301; CHECK-NEXT: [[TMP6:%.*]] = mul nsw i64 [[TMP5]], [[INDVARS_IV_NEXT]] 302; CHECK-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i32, i32* [[IN:%.*]], i64 [[TMP6]] 303; CHECK-NEXT: [[TMP7:%.*]] = load i32, i32* [[ADD_PTR]], align 4 304; CHECK-NEXT: [[TMP8:%.*]] = mul nsw i64 [[TMP1]], [[INDVARS_IV_NEXT]] 305; CHECK-NEXT: [[ADD_PTR1:%.*]] = getelementptr inbounds i32, i32* [[IN]], i64 [[TMP8]] 306; CHECK-NEXT: [[TMP9:%.*]] = load i32, i32* [[ADD_PTR1]], align 4 307; CHECK-NEXT: [[TMP10]] = add i32 [[TMP7]], [[TMP9]] 308; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 309; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_CLEANUP_LOOPEXIT]] 310; 311entry: 312 %stride = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 1 313 %0 = load i32, i32* %stride, align 4 314 %cmp17 = icmp sgt i32 %length, 1 315 br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup 316 317for.body.lr.ph: ; preds = %entry 318 %channel = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 0 319 br label %for.body 320 321for.cond.cleanup.loopexit: ; preds = %for.body 322 %1 = phi i32 [ %6, %for.body ] 323 br label %for.cond.cleanup 324 325for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 326 %2 = phi i32 [ 0, %entry ], [ %1, %for.cond.cleanup.loopexit ] 327 ret i32 %2 328 329; mul instruction below is widened instead of generating a truncate instruction for it 330; regardless if Load operand of mul is inside or outside the loop (we have both cases). 331for.body: ; preds = %for.body.lr.ph, %for.body 332 %x.018 = phi i32 [ 1, %for.body.lr.ph ], [ %add, %for.body ] 333 %add = add nuw nsw i32 %x.018, 1 334 %3 = load i32, i32* %channel, align 8 335 %mul = mul nsw i32 %3, %add 336 %idx.ext = sext i32 %mul to i64 337 %add.ptr = getelementptr inbounds i32, i32* %in, i64 %idx.ext 338 %4 = load i32, i32* %add.ptr, align 4 339 %mul1 = mul nsw i32 %0, %add 340 %idx.ext1 = sext i32 %mul1 to i64 341 %add.ptr1 = getelementptr inbounds i32, i32* %in, i64 %idx.ext1 342 %5 = load i32, i32* %add.ptr1, align 4 343 %6 = add i32 %4, %5 344 %cmp = icmp slt i32 %add, %length 345 br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit 346} 347 348 349define i32 @foo5(%struct.image* %input, i32 %length, i32* %in) { 350; CHECK-LABEL: @foo5( 351; CHECK-NEXT: entry: 352; CHECK-NEXT: [[STRIDE:%.*]] = getelementptr inbounds [[STRUCT_IMAGE:%.*]], %struct.image* [[INPUT:%.*]], i64 0, i32 1 353; CHECK-NEXT: [[TMP0:%.*]] = load i32, i32* [[STRIDE]], align 4 354; CHECK-NEXT: [[CMP17:%.*]] = icmp sgt i32 [[LENGTH:%.*]], 1 355; CHECK-NEXT: br i1 [[CMP17]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]] 356; CHECK: for.body.lr.ph: 357; CHECK-NEXT: [[CHANNEL:%.*]] = getelementptr inbounds [[STRUCT_IMAGE]], %struct.image* [[INPUT]], i64 0, i32 0 358; CHECK-NEXT: [[TMP1:%.*]] = sext i32 [[TMP0]] to i64 359; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[LENGTH]] to i64 360; CHECK-NEXT: br label [[FOR_BODY:%.*]] 361; CHECK: for.cond.cleanup.loopexit: 362; CHECK-NEXT: [[TMP2:%.*]] = phi i32 [ [[TMP10:%.*]], [[FOR_BODY]] ] 363; CHECK-NEXT: br label [[FOR_COND_CLEANUP]] 364; CHECK: for.cond.cleanup: 365; CHECK-NEXT: [[TMP3:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[TMP2]], [[FOR_COND_CLEANUP_LOOPEXIT:%.*]] ] 366; CHECK-NEXT: ret i32 [[TMP3]] 367; CHECK: for.body: 368; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ 1, [[FOR_BODY_LR_PH]] ] 369; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 370; CHECK-NEXT: [[TMP4:%.*]] = load i32, i32* [[CHANNEL]], align 8 371; CHECK-NEXT: [[TMP5:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32 372; CHECK-NEXT: [[MUL:%.*]] = mul nsw i32 [[TMP4]], [[TMP5]] 373; CHECK-NEXT: [[IDX_EXT:%.*]] = sext i32 [[MUL]] to i64 374; CHECK-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i32, i32* [[IN:%.*]], i64 [[IDX_EXT]] 375; CHECK-NEXT: [[TMP6:%.*]] = load i32, i32* [[ADD_PTR]], align 4 376; CHECK-NEXT: [[TMP7:%.*]] = mul nsw i64 [[TMP1]], [[INDVARS_IV_NEXT]] 377; CHECK-NEXT: [[ADD_PTR1:%.*]] = getelementptr inbounds i32, i32* [[IN]], i64 [[TMP7]] 378; CHECK-NEXT: [[TMP8:%.*]] = load i32, i32* [[ADD_PTR1]], align 4 379; CHECK-NEXT: [[TMP9:%.*]] = add i32 [[TMP6]], [[TMP8]] 380; CHECK-NEXT: [[TMP10]] = add i32 [[TMP9]], [[MUL]] 381; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]] 382; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_CLEANUP_LOOPEXIT]] 383; 384entry: 385 %stride = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 1 386 %0 = load i32, i32* %stride, align 4 387 %cmp17 = icmp sgt i32 %length, 1 388 br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup 389 390for.body.lr.ph: ; preds = %entry 391 %channel = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 0 392 br label %for.body 393 394for.cond.cleanup.loopexit: ; preds = %for.body 395 %1 = phi i32 [ %7, %for.body ] 396 br label %for.cond.cleanup 397 398for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 399 %2 = phi i32 [ 0, %entry ], [ %1, %for.cond.cleanup.loopexit ] 400 ret i32 %2 401 402; This example is the same as above except that the first mul is used in two places 403; and this may result in having two versions of the multiply: an i32 and i64 version. 404; In this case, keep the trucate instructions to avoid this redundancy. 405for.body: ; preds = %for.body.lr.ph, %for.body 406 %x.018 = phi i32 [ 1, %for.body.lr.ph ], [ %add, %for.body ] 407 %add = add nuw nsw i32 %x.018, 1 408 %3 = load i32, i32* %channel, align 8 409 %mul = mul nsw i32 %3, %add 410 %idx.ext = sext i32 %mul to i64 411 %add.ptr = getelementptr inbounds i32, i32* %in, i64 %idx.ext 412 %4 = load i32, i32* %add.ptr, align 4 413 %mul1 = mul nsw i32 %0, %add 414 %idx.ext1 = sext i32 %mul1 to i64 415 %add.ptr1 = getelementptr inbounds i32, i32* %in, i64 %idx.ext1 416 %5 = load i32, i32* %add.ptr1, align 4 417 %6 = add i32 %4, %5 418 %7 = add i32 %6, %mul 419 %cmp = icmp slt i32 %add, %length 420 br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit 421} 422 423define i32 @foo6(%struct.image* %input, i32 %length, i32* %in) { 424entry: 425 %stride = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 1 426 %0 = load i32, i32* %stride, align 4 427 %cmp17 = icmp sgt i32 %length, 1 428 br i1 %cmp17, label %for.body.lr.ph, label %for.cond.cleanup 429 430for.body.lr.ph: ; preds = %entry 431 %channel = getelementptr inbounds %struct.image, %struct.image* %input, i64 0, i32 0 432 br label %for.body 433 434for.cond.cleanup.loopexit: ; preds = %for.body 435 %1 = phi i32 [ %6, %for.body ] 436 br label %for.cond.cleanup 437 438for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 439 %2 = phi i32 [ 0, %entry ], [ %1, %for.cond.cleanup.loopexit ] 440 ret i32 %2 441 442; Extend foo4 so that any loop variants (%3 and %or) with mul/sub/add then extend will not 443; need a trunc instruction 444; CHECK: for.body: 445; CHECK-NOT: trunc 446; CHECK: [[TMP0:%.*]] = and i32 %length, %0 447; CHECK-NEXT: zext i32 [[TMP0]] to i64 448; CHECK: [[TMP1:%.*]] = or i32 %length, [[TMP2:%.*]] 449; CHECK-NEXT: zext i32 [[TMP1]] to i64 450for.body: ; preds = %for.body.lr.ph, %for.body 451 %x.018 = phi i32 [ 1, %for.body.lr.ph ], [ %add, %for.body ] 452 %add = add nuw nsw i32 %x.018, 1 453 %3 = and i32 %length, %0 454 %mul = mul nuw i32 %3, %add 455 %idx.ext = zext i32 %mul to i64 456 %add.ptr = getelementptr inbounds i32, i32* %in, i64 %idx.ext 457 %4 = load i32, i32* %add.ptr, align 4 458 %mul1 = mul nuw i32 %0, %add 459 %idx.ext1 = zext i32 %mul1 to i64 460 %add.ptr1 = getelementptr inbounds i32, i32* %in, i64 %idx.ext1 461 %5 = load i32, i32* %add.ptr1, align 4 462 %or = or i32 %length, %5 463 %sub.or = sub nuw i32 %or, %add 464 %or.ext = zext i32 %sub.or to i64 465 %ptr.or = getelementptr inbounds i32, i32* %in, i64 %or.ext 466 %val.or = load i32, i32* %ptr.or 467 %6 = add i32 %4, %val.or 468 %cmp = icmp ult i32 %add, %length 469 br i1 %cmp, label %for.body, label %for.cond.cleanup.loopexit 470} 471