1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -loop-versioning -S < %s | FileCheck %s -check-prefix=LV 3 4target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" 5 6; For this loop: 7; unsigned index = 0; 8; for (int i = 0; i < n; i++) { 9; A[2 * index] = A[2 * index] + B[i]; 10; index++; 11; } 12; 13; SCEV is unable to prove that A[2 * i] does not overflow. 14; 15; Analyzing the IR does not help us because the GEPs are not 16; affine AddRecExprs. However, we can turn them into AddRecExprs 17; using SCEV Predicates. 18; 19; Once we have an affine expression we need to add an additional NUSW 20; to check that the pointers don't wrap since the GEPs are not 21; inbound. 22 23; The expression for %mul_ext as analyzed by SCEV is 24; (zext i32 {0,+,2}<%for.body> to i64) 25; We have added the nusw flag to turn this expression into the SCEV expression: 26; i64 {0,+,2}<%for.body> 27 28define void @f1(i16* noalias %a, 29; LV-LABEL: @f1( 30; LV-NEXT: for.body.lver.check: 31; LV-NEXT: [[A5:%.*]] = bitcast i16* [[A:%.*]] to i8* 32; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1 33; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 34; LV-NEXT: [[TMP8:%.*]] = or i1 false, [[TMP7]] 35; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 36; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 37; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 38; LV-NEXT: [[TMP11:%.*]] = sub i64 0, [[MUL_RESULT3]] 39; LV-NEXT: [[TMP12:%.*]] = getelementptr i8, i8* [[A5]], i64 [[MUL_RESULT3]] 40; LV-NEXT: [[TMP15:%.*]] = icmp ult i8* [[TMP12]], [[A5]] 41; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW4]] 42; LV-NEXT: [[TMP18:%.*]] = or i1 [[TMP8]], [[TMP17]] 43; LV-NEXT: br i1 [[TMP18]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 44; LV: for.body.ph.lver.orig: 45; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 46; LV: for.body.lver.orig: 47; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 48; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 49; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 50; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64 51; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 52; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 53; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 54; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 55; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 56; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 57; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 58; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1 59; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 60; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 61; LV: for.body.ph: 62; LV-NEXT: br label [[FOR_BODY:%.*]] 63; LV: for.body: 64; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 65; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ] 66; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 67; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64 68; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 69; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 70; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 71; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 72; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 73; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 74; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 75; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1 76; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 77; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT3:%.*]], label [[FOR_BODY]] 78; LV: for.end.loopexit: 79; LV-NEXT: br label [[FOR_END:%.*]] 80; LV: for.end.loopexit3: 81; LV-NEXT: br label [[FOR_END]] 82; LV: for.end: 83; LV-NEXT: ret void 84; 85 i16* noalias %b, i64 %N) { 86entry: 87 br label %for.body 88 89for.body: ; preds = %for.body, %entry 90 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 91 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ] 92 93 %mul = mul i32 %ind1, 2 94 %mul_ext = zext i32 %mul to i64 95 96 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 97 %loadA = load i16, i16* %arrayidxA, align 2 98 99 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 100 %loadB = load i16, i16* %arrayidxB, align 2 101 102 %add = mul i16 %loadA, %loadB 103 104 store i16 %add, i16* %arrayidxA, align 2 105 106 %inc = add nuw nsw i64 %ind, 1 107 %inc1 = add i32 %ind1, 1 108 109 %exitcond = icmp eq i64 %inc, %N 110 br i1 %exitcond, label %for.end, label %for.body 111 112for.end: ; preds = %for.body 113 ret void 114} 115 116; For this loop: 117; unsigned index = n; 118; for (int i = 0; i < n; i++) { 119; A[2 * index] = A[2 * index] + B[i]; 120; index--; 121; } 122; 123; the SCEV expression for 2 * index is not an AddRecExpr 124; (and implictly not affine). However, we are able to make assumptions 125; that will turn the expression into an affine one and continue the 126; analysis. 127; 128; Once we have an affine expression we need to add an additional NUSW 129; to check that the pointers don't wrap since the GEPs are not 130; inbounds. 131; 132; This loop has a negative stride for A, and the nusw flag is required in 133; order to properly extend the increment from i32 -4 to i64 -4. 134 135; The expression for %mul_ext as analyzed by SCEV is 136; (zext i32 {(2 * (trunc i64 %N to i32)),+,-2}<%for.body> to i64) 137; We have added the nusw flag to turn this expression into the following SCEV: 138; i64 {zext i32 (2 * (trunc i64 %N to i32)) to i64,+,-2}<%for.body> 139 140define void @f2(i16* noalias %a, 141; LV-LABEL: @f2( 142; LV-NEXT: for.body.lver.check: 143; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 144; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 145; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 146; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 147; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 148; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 149; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 150; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 151; LV-NEXT: [[TMP5:%.*]] = icmp ugt i32 [[TMP4]], [[TMP1]] 152; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 153; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 154; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 155; LV-NEXT: [[TMP12:%.*]] = trunc i64 [[N]] to i31 156; LV-NEXT: [[TMP13:%.*]] = zext i31 [[TMP12]] to i64 157; LV-NEXT: [[TMP14:%.*]] = shl nuw nsw i64 [[TMP13]], 1 158; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP14]] 159; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 160; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 161; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 162; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 163; LV-NEXT: [[TMP15:%.*]] = sub i64 0, [[MUL_RESULT3]] 164; LV-NEXT: [[TMP17:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP15]] 165; LV-NEXT: [[TMP18:%.*]] = icmp ugt i8* [[TMP17]], [[SCEVGEP5]] 166; LV-NEXT: [[TMP21:%.*]] = or i1 [[TMP18]], [[MUL_OVERFLOW4]] 167; LV-NEXT: [[TMP22:%.*]] = or i1 [[TMP10]], [[TMP21]] 168; LV-NEXT: br i1 [[TMP22]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 169; LV: for.body.ph.lver.orig: 170; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 171; LV: for.body.lver.orig: 172; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 173; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 174; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 175; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64 176; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 177; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 178; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 179; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 180; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 181; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 182; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 183; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 184; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 185; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 186; LV: for.body.ph: 187; LV-NEXT: br label [[FOR_BODY:%.*]] 188; LV: for.body: 189; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 190; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 191; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 192; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64 193; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 194; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 195; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 196; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 197; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 198; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 199; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 200; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 201; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 202; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 203; LV: for.end.loopexit: 204; LV-NEXT: br label [[FOR_END:%.*]] 205; LV: for.end.loopexit6: 206; LV-NEXT: br label [[FOR_END]] 207; LV: for.end: 208; LV-NEXT: ret void 209; 210 i16* noalias %b, i64 %N) { 211entry: 212 %TruncN = trunc i64 %N to i32 213 br label %for.body 214 215for.body: ; preds = %for.body, %entry 216 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 217 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 218 219 %mul = mul i32 %ind1, 2 220 %mul_ext = zext i32 %mul to i64 221 222 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 223 %loadA = load i16, i16* %arrayidxA, align 2 224 225 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 226 %loadB = load i16, i16* %arrayidxB, align 2 227 228 %add = mul i16 %loadA, %loadB 229 230 store i16 %add, i16* %arrayidxA, align 2 231 232 %inc = add nuw nsw i64 %ind, 1 233 %dec = sub i32 %ind1, 1 234 235 %exitcond = icmp eq i64 %inc, %N 236 br i1 %exitcond, label %for.end, label %for.body 237 238for.end: ; preds = %for.body 239 ret void 240} 241 242; We replicate the tests above, but this time sign extend 2 * index instead 243; of zero extending it. 244 245; The expression for %mul_ext as analyzed by SCEV is 246; i64 (sext i32 {0,+,2}<%for.body> to i64) 247; We have added the nssw flag to turn this expression into the following SCEV: 248; i64 {0,+,2}<%for.body> 249 250define void @f3(i16* noalias %a, 251; LV-LABEL: @f3( 252; LV-NEXT: for.body.lver.check: 253; LV-NEXT: [[A5:%.*]] = bitcast i16* [[A:%.*]] to i8* 254; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1 255; LV-NEXT: [[TMP1:%.*]] = trunc i64 [[TMP0]] to i32 256; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP1]]) 257; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 258; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 259; LV-NEXT: [[TMP2:%.*]] = add i32 0, [[MUL_RESULT]] 260; LV-NEXT: [[TMP5:%.*]] = icmp slt i32 [[TMP2]], 0 261; LV-NEXT: [[TMP8:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 262; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 263; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[TMP7]] 264; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 265; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 266; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 267; LV-NEXT: [[TMP11:%.*]] = sub i64 0, [[MUL_RESULT3]] 268; LV-NEXT: [[TMP12:%.*]] = getelementptr i8, i8* [[A5]], i64 [[MUL_RESULT3]] 269; LV-NEXT: [[TMP15:%.*]] = icmp ult i8* [[TMP12]], [[A5]] 270; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW4]] 271; LV-NEXT: [[TMP18:%.*]] = or i1 [[TMP9]], [[TMP17]] 272; LV-NEXT: br i1 [[TMP18]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 273; LV: for.body.ph.lver.orig: 274; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 275; LV: for.body.lver.orig: 276; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 277; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 278; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 279; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64 280; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 281; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 282; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 283; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 284; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 285; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 286; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 287; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1 288; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 289; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 290; LV: for.body.ph: 291; LV-NEXT: br label [[FOR_BODY:%.*]] 292; LV: for.body: 293; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 294; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ] 295; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 296; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64 297; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 298; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 299; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 300; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 301; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 302; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 303; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 304; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1 305; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 306; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 307; LV: for.end.loopexit: 308; LV-NEXT: br label [[FOR_END:%.*]] 309; LV: for.end.loopexit6: 310; LV-NEXT: br label [[FOR_END]] 311; LV: for.end: 312; LV-NEXT: ret void 313; 314 i16* noalias %b, i64 %N) { 315entry: 316 br label %for.body 317 318for.body: ; preds = %for.body, %entry 319 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 320 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ] 321 322 %mul = mul i32 %ind1, 2 323 %mul_ext = sext i32 %mul to i64 324 325 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 326 %loadA = load i16, i16* %arrayidxA, align 2 327 328 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 329 %loadB = load i16, i16* %arrayidxB, align 2 330 331 %add = mul i16 %loadA, %loadB 332 333 store i16 %add, i16* %arrayidxA, align 2 334 335 %inc = add nuw nsw i64 %ind, 1 336 %inc1 = add i32 %ind1, 1 337 338 %exitcond = icmp eq i64 %inc, %N 339 br i1 %exitcond, label %for.end, label %for.body 340 341for.end: ; preds = %for.body 342 ret void 343} 344 345define void @f4(i16* noalias %a, 346; LV-LABEL: @f4( 347; LV-NEXT: for.body.lver.check: 348; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 349; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 350; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 351; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 352; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 353; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 354; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 355; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 356; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]] 357; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 358; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 359; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 360; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64 361; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP12]] 362; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 363; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 364; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 365; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 366; LV-NEXT: [[TMP13:%.*]] = sub i64 0, [[MUL_RESULT3]] 367; LV-NEXT: [[TMP15:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP13]] 368; LV-NEXT: [[TMP16:%.*]] = icmp ugt i8* [[TMP15]], [[SCEVGEP5]] 369; LV-NEXT: [[TMP19:%.*]] = or i1 [[TMP16]], [[MUL_OVERFLOW4]] 370; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP10]], [[TMP19]] 371; LV-NEXT: br i1 [[TMP20]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 372; LV: for.body.ph.lver.orig: 373; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 374; LV: for.body.lver.orig: 375; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 376; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 377; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 378; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64 379; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 380; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 381; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 382; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 383; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 384; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 385; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 386; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 387; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 388; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 389; LV: for.body.ph: 390; LV-NEXT: br label [[FOR_BODY:%.*]] 391; LV: for.body: 392; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 393; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 394; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 395; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64 396; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 397; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 398; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 399; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 400; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 401; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 402; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 403; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 404; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 405; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 406; LV: for.end.loopexit: 407; LV-NEXT: br label [[FOR_END:%.*]] 408; LV: for.end.loopexit6: 409; LV-NEXT: br label [[FOR_END]] 410; LV: for.end: 411; LV-NEXT: ret void 412; 413 i16* noalias %b, i64 %N) { 414entry: 415 %TruncN = trunc i64 %N to i32 416 br label %for.body 417 418for.body: ; preds = %for.body, %entry 419 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 420 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 421 422 %mul = mul i32 %ind1, 2 423 %mul_ext = sext i32 %mul to i64 424 425 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 426 %loadA = load i16, i16* %arrayidxA, align 2 427 428 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 429 %loadB = load i16, i16* %arrayidxB, align 2 430 431 %add = mul i16 %loadA, %loadB 432 433 store i16 %add, i16* %arrayidxA, align 2 434 435 %inc = add nuw nsw i64 %ind, 1 436 %dec = sub i32 %ind1, 1 437 438 %exitcond = icmp eq i64 %inc, %N 439 br i1 %exitcond, label %for.end, label %for.body 440 441for.end: ; preds = %for.body 442 ret void 443} 444 445; The following function is similar to the one above, but has the GEP 446; to pointer %A inbounds. The index %mul doesn't have the nsw flag. 447; This means that the SCEV expression for %mul can wrap and we need 448; a SCEV predicate to continue analysis. 449; 450; We can still analyze this by adding the required no wrap SCEV predicates. 451 452define void @f5(i16* noalias %a, 453; LV-LABEL: @f5( 454; LV-NEXT: for.body.lver.check: 455; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 456; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 457; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 458; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 459; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 460; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 461; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 462; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 463; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]] 464; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 465; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 466; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 467; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64 468; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP12]] 469; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 470; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 471; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 472; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 473; LV-NEXT: [[TMP13:%.*]] = sub i64 0, [[MUL_RESULT3]] 474; LV-NEXT: [[TMP15:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP13]] 475; LV-NEXT: [[TMP16:%.*]] = icmp ugt i8* [[TMP15]], [[SCEVGEP5]] 476; LV-NEXT: [[TMP19:%.*]] = or i1 [[TMP16]], [[MUL_OVERFLOW4]] 477; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP10]], [[TMP19]] 478; LV-NEXT: br i1 [[TMP20]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 479; LV: for.body.ph.lver.orig: 480; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 481; LV: for.body.lver.orig: 482; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 483; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 484; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 485; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL_LVER_ORIG]] 486; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 487; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 488; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 489; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 490; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 491; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 492; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 493; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 494; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 495; LV: for.body.ph: 496; LV-NEXT: br label [[FOR_BODY:%.*]] 497; LV: for.body: 498; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 499; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 500; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 501; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL]] 502; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 503; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr inbounds i16, i16* [[B]], i64 [[IND]] 504; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 505; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 506; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 507; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 508; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 509; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 510; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 511; LV: for.end.loopexit: 512; LV-NEXT: br label [[FOR_END:%.*]] 513; LV: for.end.loopexit6: 514; LV-NEXT: br label [[FOR_END]] 515; LV: for.end: 516; LV-NEXT: ret void 517; 518 i16* noalias %b, i64 %N) { 519entry: 520 %TruncN = trunc i64 %N to i32 521 br label %for.body 522 523for.body: ; preds = %for.body, %entry 524 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 525 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 526 527 %mul = mul i32 %ind1, 2 528 529 %arrayidxA = getelementptr inbounds i16, i16* %a, i32 %mul 530 %loadA = load i16, i16* %arrayidxA, align 2 531 532 %arrayidxB = getelementptr inbounds i16, i16* %b, i64 %ind 533 %loadB = load i16, i16* %arrayidxB, align 2 534 535 %add = mul i16 %loadA, %loadB 536 537 store i16 %add, i16* %arrayidxA, align 2 538 539 %inc = add nuw nsw i64 %ind, 1 540 %dec = sub i32 %ind1, 1 541 542 %exitcond = icmp eq i64 %inc, %N 543 br i1 %exitcond, label %for.end, label %for.body 544 545for.end: ; preds = %for.body 546 ret void 547} 548