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: [[TMP1:%.*]] = trunc i64 [[TMP0]] to i32 34; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP1]]) 35; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 36; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 37; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 38; LV-NEXT: [[TMP8:%.*]] = or i1 false, [[TMP7]] 39; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 40; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 41; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 42; LV-NEXT: [[TMP11:%.*]] = sub i64 0, [[MUL_RESULT3]] 43; LV-NEXT: [[TMP12:%.*]] = getelementptr i8, i8* [[A5]], i64 [[MUL_RESULT3]] 44; LV-NEXT: [[TMP15:%.*]] = icmp ult i8* [[TMP12]], [[A5]] 45; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW4]] 46; LV-NEXT: [[TMP18:%.*]] = or i1 [[TMP8]], [[TMP17]] 47; LV-NEXT: br i1 [[TMP18]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 48; LV: for.body.ph.lver.orig: 49; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 50; LV: for.body.lver.orig: 51; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 52; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 53; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 54; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64 55; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 56; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 57; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 58; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 59; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 60; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 61; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 62; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1 63; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 64; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 65; LV: for.body.ph: 66; LV-NEXT: br label [[FOR_BODY:%.*]] 67; LV: for.body: 68; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 69; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ] 70; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 71; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64 72; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 73; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 74; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 75; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 76; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 77; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 78; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 79; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1 80; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 81; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 82; LV: for.end.loopexit: 83; LV-NEXT: br label [[FOR_END:%.*]] 84; LV: for.end.loopexit6: 85; LV-NEXT: br label [[FOR_END]] 86; LV: for.end: 87; LV-NEXT: ret void 88; 89 i16* noalias %b, i64 %N) { 90entry: 91 br label %for.body 92 93for.body: ; preds = %for.body, %entry 94 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 95 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ] 96 97 %mul = mul i32 %ind1, 2 98 %mul_ext = zext i32 %mul to i64 99 100 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 101 %loadA = load i16, i16* %arrayidxA, align 2 102 103 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 104 %loadB = load i16, i16* %arrayidxB, align 2 105 106 %add = mul i16 %loadA, %loadB 107 108 store i16 %add, i16* %arrayidxA, align 2 109 110 %inc = add nuw nsw i64 %ind, 1 111 %inc1 = add i32 %ind1, 1 112 113 %exitcond = icmp eq i64 %inc, %N 114 br i1 %exitcond, label %for.end, label %for.body 115 116for.end: ; preds = %for.body 117 ret void 118} 119 120; For this loop: 121; unsigned index = n; 122; for (int i = 0; i < n; i++) { 123; A[2 * index] = A[2 * index] + B[i]; 124; index--; 125; } 126; 127; the SCEV expression for 2 * index is not an AddRecExpr 128; (and implictly not affine). However, we are able to make assumptions 129; that will turn the expression into an affine one and continue the 130; analysis. 131; 132; Once we have an affine expression we need to add an additional NUSW 133; to check that the pointers don't wrap since the GEPs are not 134; inbounds. 135; 136; This loop has a negative stride for A, and the nusw flag is required in 137; order to properly extend the increment from i32 -4 to i64 -4. 138 139; The expression for %mul_ext as analyzed by SCEV is 140; (zext i32 {(2 * (trunc i64 %N to i32)),+,-2}<%for.body> to i64) 141; We have added the nusw flag to turn this expression into the following SCEV: 142; i64 {zext i32 (2 * (trunc i64 %N to i32)) to i64,+,-2}<%for.body> 143 144define void @f2(i16* noalias %a, 145; LV-LABEL: @f2( 146; LV-NEXT: for.body.lver.check: 147; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 148; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 149; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 150; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 151; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 152; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 153; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 154; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 155; LV-NEXT: [[TMP5:%.*]] = icmp ugt i32 [[TMP4]], [[TMP1]] 156; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 157; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 158; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 159; LV-NEXT: [[TMP12:%.*]] = trunc i64 [[N]] to i31 160; LV-NEXT: [[TMP13:%.*]] = zext i31 [[TMP12]] to i64 161; LV-NEXT: [[TMP14:%.*]] = shl nuw nsw i64 [[TMP13]], 1 162; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP14]] 163; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 164; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 165; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 166; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 167; LV-NEXT: [[TMP15:%.*]] = sub i64 0, [[MUL_RESULT3]] 168; LV-NEXT: [[TMP17:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP15]] 169; LV-NEXT: [[TMP18:%.*]] = icmp ugt i8* [[TMP17]], [[SCEVGEP5]] 170; LV-NEXT: [[TMP21:%.*]] = or i1 [[TMP18]], [[MUL_OVERFLOW4]] 171; LV-NEXT: [[TMP22:%.*]] = or i1 [[TMP10]], [[TMP21]] 172; LV-NEXT: br i1 [[TMP22]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 173; LV: for.body.ph.lver.orig: 174; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 175; LV: for.body.lver.orig: 176; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 177; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 178; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 179; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64 180; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 181; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 182; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 183; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 184; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 185; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 186; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 187; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 188; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 189; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 190; LV: for.body.ph: 191; LV-NEXT: br label [[FOR_BODY:%.*]] 192; LV: for.body: 193; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 194; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 195; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 196; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64 197; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 198; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 199; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 200; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 201; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 202; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 203; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 204; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 205; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 206; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 207; LV: for.end.loopexit: 208; LV-NEXT: br label [[FOR_END:%.*]] 209; LV: for.end.loopexit6: 210; LV-NEXT: br label [[FOR_END]] 211; LV: for.end: 212; LV-NEXT: ret void 213; 214 i16* noalias %b, i64 %N) { 215entry: 216 %TruncN = trunc i64 %N to i32 217 br label %for.body 218 219for.body: ; preds = %for.body, %entry 220 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 221 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 222 223 %mul = mul i32 %ind1, 2 224 %mul_ext = zext i32 %mul to i64 225 226 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 227 %loadA = load i16, i16* %arrayidxA, align 2 228 229 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 230 %loadB = load i16, i16* %arrayidxB, align 2 231 232 %add = mul i16 %loadA, %loadB 233 234 store i16 %add, i16* %arrayidxA, align 2 235 236 %inc = add nuw nsw i64 %ind, 1 237 %dec = sub i32 %ind1, 1 238 239 %exitcond = icmp eq i64 %inc, %N 240 br i1 %exitcond, label %for.end, label %for.body 241 242for.end: ; preds = %for.body 243 ret void 244} 245 246; We replicate the tests above, but this time sign extend 2 * index instead 247; of zero extending it. 248 249; The expression for %mul_ext as analyzed by SCEV is 250; i64 (sext i32 {0,+,2}<%for.body> to i64) 251; We have added the nssw flag to turn this expression into the following SCEV: 252; i64 {0,+,2}<%for.body> 253 254define void @f3(i16* noalias %a, 255; LV-LABEL: @f3( 256; LV-NEXT: for.body.lver.check: 257; LV-NEXT: [[A5:%.*]] = bitcast i16* [[A:%.*]] to i8* 258; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1 259; LV-NEXT: [[TMP1:%.*]] = trunc i64 [[TMP0]] to i32 260; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP1]]) 261; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 262; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 263; LV-NEXT: [[TMP2:%.*]] = add i32 0, [[MUL_RESULT]] 264; LV-NEXT: [[TMP5:%.*]] = icmp slt i32 [[TMP2]], 0 265; LV-NEXT: [[TMP8:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 266; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 267; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[TMP7]] 268; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 269; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 270; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 271; LV-NEXT: [[TMP11:%.*]] = sub i64 0, [[MUL_RESULT3]] 272; LV-NEXT: [[TMP12:%.*]] = getelementptr i8, i8* [[A5]], i64 [[MUL_RESULT3]] 273; LV-NEXT: [[TMP15:%.*]] = icmp ult i8* [[TMP12]], [[A5]] 274; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW4]] 275; LV-NEXT: [[TMP18:%.*]] = or i1 [[TMP9]], [[TMP17]] 276; LV-NEXT: br i1 [[TMP18]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 277; LV: for.body.ph.lver.orig: 278; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 279; LV: for.body.lver.orig: 280; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 281; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 282; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 283; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64 284; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 285; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 286; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 287; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 288; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 289; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 290; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 291; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1 292; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 293; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 294; LV: for.body.ph: 295; LV-NEXT: br label [[FOR_BODY:%.*]] 296; LV: for.body: 297; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 298; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ] 299; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 300; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64 301; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 302; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 303; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 304; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 305; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 306; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 307; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 308; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1 309; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 310; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 311; LV: for.end.loopexit: 312; LV-NEXT: br label [[FOR_END:%.*]] 313; LV: for.end.loopexit6: 314; LV-NEXT: br label [[FOR_END]] 315; LV: for.end: 316; LV-NEXT: ret void 317; 318 i16* noalias %b, i64 %N) { 319entry: 320 br label %for.body 321 322for.body: ; preds = %for.body, %entry 323 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 324 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ] 325 326 %mul = mul i32 %ind1, 2 327 %mul_ext = sext i32 %mul to i64 328 329 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 330 %loadA = load i16, i16* %arrayidxA, align 2 331 332 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 333 %loadB = load i16, i16* %arrayidxB, align 2 334 335 %add = mul i16 %loadA, %loadB 336 337 store i16 %add, i16* %arrayidxA, align 2 338 339 %inc = add nuw nsw i64 %ind, 1 340 %inc1 = add i32 %ind1, 1 341 342 %exitcond = icmp eq i64 %inc, %N 343 br i1 %exitcond, label %for.end, label %for.body 344 345for.end: ; preds = %for.body 346 ret void 347} 348 349define void @f4(i16* noalias %a, 350; LV-LABEL: @f4( 351; LV-NEXT: for.body.lver.check: 352; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 353; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 354; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 355; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 356; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 357; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 358; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 359; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 360; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]] 361; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 362; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 363; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 364; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64 365; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP12]] 366; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 367; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 368; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 369; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 370; LV-NEXT: [[TMP13:%.*]] = sub i64 0, [[MUL_RESULT3]] 371; LV-NEXT: [[TMP15:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP13]] 372; LV-NEXT: [[TMP16:%.*]] = icmp ugt i8* [[TMP15]], [[SCEVGEP5]] 373; LV-NEXT: [[TMP19:%.*]] = or i1 [[TMP16]], [[MUL_OVERFLOW4]] 374; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP10]], [[TMP19]] 375; LV-NEXT: br i1 [[TMP20]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 376; LV: for.body.ph.lver.orig: 377; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 378; LV: for.body.lver.orig: 379; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 380; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 381; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 382; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64 383; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]] 384; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 385; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 386; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 387; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 388; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 389; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 390; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 391; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 392; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 393; LV: for.body.ph: 394; LV-NEXT: br label [[FOR_BODY:%.*]] 395; LV: for.body: 396; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 397; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 398; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 399; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64 400; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]] 401; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 402; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]] 403; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 404; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 405; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 406; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 407; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 408; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 409; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 410; LV: for.end.loopexit: 411; LV-NEXT: br label [[FOR_END:%.*]] 412; LV: for.end.loopexit6: 413; LV-NEXT: br label [[FOR_END]] 414; LV: for.end: 415; LV-NEXT: ret void 416; 417 i16* noalias %b, i64 %N) { 418entry: 419 %TruncN = trunc i64 %N to i32 420 br label %for.body 421 422for.body: ; preds = %for.body, %entry 423 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 424 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 425 426 %mul = mul i32 %ind1, 2 427 %mul_ext = sext i32 %mul to i64 428 429 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext 430 %loadA = load i16, i16* %arrayidxA, align 2 431 432 %arrayidxB = getelementptr i16, i16* %b, i64 %ind 433 %loadB = load i16, i16* %arrayidxB, align 2 434 435 %add = mul i16 %loadA, %loadB 436 437 store i16 %add, i16* %arrayidxA, align 2 438 439 %inc = add nuw nsw i64 %ind, 1 440 %dec = sub i32 %ind1, 1 441 442 %exitcond = icmp eq i64 %inc, %N 443 br i1 %exitcond, label %for.end, label %for.body 444 445for.end: ; preds = %for.body 446 ret void 447} 448 449; The following function is similar to the one above, but has the GEP 450; to pointer %A inbounds. The index %mul doesn't have the nsw flag. 451; This means that the SCEV expression for %mul can wrap and we need 452; a SCEV predicate to continue analysis. 453; 454; We can still analyze this by adding the required no wrap SCEV predicates. 455 456define void @f5(i16* noalias %a, 457; LV-LABEL: @f5( 458; LV-NEXT: for.body.lver.check: 459; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32 460; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1 461; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1 462; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32 463; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]]) 464; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0 465; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1 466; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]] 467; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]] 468; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP5]], [[MUL_OVERFLOW]] 469; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295 470; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP8]] 471; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64 472; LV-NEXT: [[SCEVGEP:%.*]] = getelementptr i16, i16* [[A:%.*]], i64 [[TMP12]] 473; LV-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]]) 474; LV-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0 475; LV-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1 476; LV-NEXT: [[SCEVGEP5:%.*]] = bitcast i16* [[SCEVGEP]] to i8* 477; LV-NEXT: [[TMP13:%.*]] = sub i64 0, [[MUL_RESULT3]] 478; LV-NEXT: [[TMP15:%.*]] = getelementptr i8, i8* [[SCEVGEP5]], i64 [[TMP13]] 479; LV-NEXT: [[TMP16:%.*]] = icmp ugt i8* [[TMP15]], [[SCEVGEP5]] 480; LV-NEXT: [[TMP19:%.*]] = or i1 [[TMP16]], [[MUL_OVERFLOW4]] 481; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP10]], [[TMP19]] 482; LV-NEXT: br i1 [[TMP20]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]] 483; LV: for.body.ph.lver.orig: 484; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]] 485; LV: for.body.lver.orig: 486; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 487; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ] 488; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2 489; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL_LVER_ORIG]] 490; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2 491; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]] 492; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2 493; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]] 494; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2 495; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1 496; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1 497; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]] 498; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]] 499; LV: for.body.ph: 500; LV-NEXT: br label [[FOR_BODY:%.*]] 501; LV: for.body: 502; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ] 503; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ] 504; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2 505; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL]] 506; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2 507; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr inbounds i16, i16* [[B]], i64 [[IND]] 508; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2 509; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]] 510; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2 511; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1 512; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1 513; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]] 514; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]] 515; LV: for.end.loopexit: 516; LV-NEXT: br label [[FOR_END:%.*]] 517; LV: for.end.loopexit6: 518; LV-NEXT: br label [[FOR_END]] 519; LV: for.end: 520; LV-NEXT: ret void 521; 522 i16* noalias %b, i64 %N) { 523entry: 524 %TruncN = trunc i64 %N to i32 525 br label %for.body 526 527for.body: ; preds = %for.body, %entry 528 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ] 529 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ] 530 531 %mul = mul i32 %ind1, 2 532 533 %arrayidxA = getelementptr inbounds i16, i16* %a, i32 %mul 534 %loadA = load i16, i16* %arrayidxA, align 2 535 536 %arrayidxB = getelementptr inbounds i16, i16* %b, i64 %ind 537 %loadB = load i16, i16* %arrayidxB, align 2 538 539 %add = mul i16 %loadA, %loadB 540 541 store i16 %add, i16* %arrayidxA, align 2 542 543 %inc = add nuw nsw i64 %ind, 1 544 %dec = sub i32 %ind1, 1 545 546 %exitcond = icmp eq i64 %inc, %N 547 br i1 %exitcond, label %for.end, label %for.body 548 549for.end: ; preds = %for.body 550 ret void 551} 552