1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt %s -S -loop-vectorize -force-vector-interleave=2 | FileCheck %s 3 4; Demonstrate a case where we unroll a loop, but don't vectorize it. 5; The original loop runs stores in the latch block on iterations 0 to 1022, 6; and exits when %indvars.iv = 1023. (That is, it actually runs the stores 7; for an odd number of iterations.) If we unroll by two in the "vector.body" 8; loop, we must exit to the epilogue on iteration with %indvars.iv = 1022 to 9; avoid an out of bounds access. 10 11define void @test(double* %data) { 12; CHECK-LABEL: @test( 13; CHECK-NEXT: entry: 14; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 15; CHECK: vector.ph: 16; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 17; CHECK: vector.body: 18; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 19; CHECK-NEXT: [[INDUCTION:%.*]] = add i64 [[INDEX]], 0 20; CHECK-NEXT: [[INDUCTION1:%.*]] = add i64 [[INDEX]], 1 21; CHECK-NEXT: [[TMP0:%.*]] = shl nuw nsw i64 [[INDUCTION]], 1 22; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[INDUCTION1]], 1 23; CHECK-NEXT: [[TMP2:%.*]] = or i64 [[TMP0]], 1 24; CHECK-NEXT: [[TMP3:%.*]] = or i64 [[TMP1]], 1 25; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds double, double* [[DATA:%.*]], i64 [[TMP2]] 26; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds double, double* [[DATA]], i64 [[TMP3]] 27; CHECK-NEXT: [[TMP6:%.*]] = load double, double* [[TMP4]], align 8 28; CHECK-NEXT: [[TMP7:%.*]] = load double, double* [[TMP5]], align 8 29; CHECK-NEXT: [[TMP8:%.*]] = fneg double [[TMP6]] 30; CHECK-NEXT: [[TMP9:%.*]] = fneg double [[TMP7]] 31; CHECK-NEXT: store double [[TMP8]], double* [[TMP4]], align 8 32; CHECK-NEXT: store double [[TMP9]], double* [[TMP5]], align 8 33; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2 34; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1022 35; CHECK-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 36; CHECK: middle.block: 37; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 1024, 1022 38; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]] 39; CHECK: scalar.ph: 40; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1022, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 41; CHECK-NEXT: br label [[FOR_BODY:%.*]] 42; CHECK: for.body: 43; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_LATCH:%.*]] ] 44; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 45; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 1024 46; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_LATCH]] 47; CHECK: for.latch: 48; CHECK-NEXT: [[T15:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1 49; CHECK-NEXT: [[T16:%.*]] = or i64 [[T15]], 1 50; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, double* [[DATA]], i64 [[T16]] 51; CHECK-NEXT: [[T17:%.*]] = load double, double* [[ARRAYIDX]], align 8 52; CHECK-NEXT: [[FNEG:%.*]] = fneg double [[T17]] 53; CHECK-NEXT: store double [[FNEG]], double* [[ARRAYIDX]], align 8 54; CHECK-NEXT: br label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]] 55; CHECK: for.end: 56; CHECK-NEXT: ret void 57; 58entry: 59 br label %for.body 60 61for.body: 62 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.latch ] 63 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 64 %exitcond.not = icmp eq i64 %indvars.iv.next, 1024 65 br i1 %exitcond.not, label %for.end, label %for.latch 66 67for.latch: 68 %t15 = shl nuw nsw i64 %indvars.iv, 1 69 %t16 = or i64 %t15, 1 70 %arrayidx = getelementptr inbounds double, double* %data, i64 %t16 71 %t17 = load double, double* %arrayidx, align 8 72 %fneg = fneg double %t17 73 store double %fneg, double* %arrayidx, align 8 74 br label %for.body 75 76for.end: 77 ret void 78} 79