1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -S -loop-vectorize -mtriple=x86_64-apple-darwin %s | FileCheck %s --check-prefixes=CHECK,SSE 3; RUN: opt -S -loop-vectorize -mtriple=x86_64-apple-darwin -mattr=+avx %s | FileCheck %s --check-prefixes=CHECK,AVX 4 5; Two mostly identical functions. The only difference is the presence of 6; fast-math flags on the second. The loop is a pretty simple reduction: 7 8; for (int i = 0; i < 32; ++i) 9; if (arr[i] != 42) 10; tot += arr[i]; 11 12define double @sumIfScalar(double* nocapture readonly %arr) { 13; CHECK-LABEL: @sumIfScalar( 14; CHECK-NEXT: entry: 15; CHECK-NEXT: br label [[LOOP:%.*]] 16; CHECK: loop: 17; CHECK-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[I_NEXT:%.*]], [[NEXT_ITER:%.*]] ] 18; CHECK-NEXT: [[TOT:%.*]] = phi double [ 0.000000e+00, [[ENTRY]] ], [ [[TOT_NEXT:%.*]], [[NEXT_ITER]] ] 19; CHECK-NEXT: [[ADDR:%.*]] = getelementptr double, double* [[ARR:%.*]], i32 [[I]] 20; CHECK-NEXT: [[NEXTVAL:%.*]] = load double, double* [[ADDR]], align 8 21; CHECK-NEXT: [[TST:%.*]] = fcmp une double [[NEXTVAL]], 4.200000e+01 22; CHECK-NEXT: br i1 [[TST]], label [[DO_ADD:%.*]], label [[NO_ADD:%.*]] 23; CHECK: do.add: 24; CHECK-NEXT: [[TOT_NEW:%.*]] = fadd double [[TOT]], [[NEXTVAL]] 25; CHECK-NEXT: br label [[NEXT_ITER]] 26; CHECK: no.add: 27; CHECK-NEXT: br label [[NEXT_ITER]] 28; CHECK: next.iter: 29; CHECK-NEXT: [[TOT_NEXT]] = phi double [ [[TOT]], [[NO_ADD]] ], [ [[TOT_NEW]], [[DO_ADD]] ] 30; CHECK-NEXT: [[I_NEXT]] = add i32 [[I]], 1 31; CHECK-NEXT: [[AGAIN:%.*]] = icmp ult i32 [[I_NEXT]], 32 32; CHECK-NEXT: br i1 [[AGAIN]], label [[LOOP]], label [[DONE:%.*]] 33; CHECK: done: 34; CHECK-NEXT: [[TOT_NEXT_LCSSA:%.*]] = phi double [ [[TOT_NEXT]], [[NEXT_ITER]] ] 35; CHECK-NEXT: ret double [[TOT_NEXT_LCSSA]] 36; 37entry: 38 br label %loop 39 40loop: 41 %i = phi i32 [0, %entry], [%i.next, %next.iter] 42 %tot = phi double [0.0, %entry], [%tot.next, %next.iter] 43 44 %addr = getelementptr double, double* %arr, i32 %i 45 %nextval = load double, double* %addr 46 47 %tst = fcmp une double %nextval, 42.0 48 br i1 %tst, label %do.add, label %no.add 49 50do.add: 51 %tot.new = fadd double %tot, %nextval 52 br label %next.iter 53 54no.add: 55 br label %next.iter 56 57next.iter: 58 %tot.next = phi double [%tot, %no.add], [%tot.new, %do.add] 59 %i.next = add i32 %i, 1 60 %again = icmp ult i32 %i.next, 32 61 br i1 %again, label %loop, label %done 62 63done: 64 ret double %tot.next 65} 66 67define double @sumIfVector(double* nocapture readonly %arr) { 68; SSE-LABEL: @sumIfVector( 69; SSE-NEXT: entry: 70; SSE-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 71; SSE: vector.ph: 72; SSE-NEXT: br label [[VECTOR_BODY:%.*]] 73; SSE: vector.body: 74; SSE-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 75; SSE-NEXT: [[VEC_PHI:%.*]] = phi <2 x double> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PREDPHI:%.*]], [[VECTOR_BODY]] ] 76; SSE-NEXT: [[TMP0:%.*]] = add i32 [[INDEX]], 0 77; SSE-NEXT: [[TMP1:%.*]] = getelementptr double, double* [[ARR:%.*]], i32 [[TMP0]] 78; SSE-NEXT: [[TMP2:%.*]] = getelementptr double, double* [[TMP1]], i32 0 79; SSE-NEXT: [[TMP3:%.*]] = bitcast double* [[TMP2]] to <2 x double>* 80; SSE-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, <2 x double>* [[TMP3]], align 8 81; SSE-NEXT: [[TMP4:%.*]] = fcmp fast une <2 x double> [[WIDE_LOAD]], <double 4.200000e+01, double 4.200000e+01> 82; SSE-NEXT: [[TMP5:%.*]] = fadd fast <2 x double> [[VEC_PHI]], [[WIDE_LOAD]] 83; SSE-NEXT: [[TMP6:%.*]] = xor <2 x i1> [[TMP4]], <i1 true, i1 true> 84; SSE-NEXT: [[PREDPHI]] = select <2 x i1> [[TMP4]], <2 x double> [[TMP5]], <2 x double> [[VEC_PHI]] 85; SSE-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 86; SSE-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], 32 87; SSE-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 88; SSE: middle.block: 89; SSE-NEXT: [[TMP8:%.*]] = call fast double @llvm.vector.reduce.fadd.v2f64(double -0.000000e+00, <2 x double> [[PREDPHI]]) 90; SSE-NEXT: [[CMP_N:%.*]] = icmp eq i32 32, 32 91; SSE-NEXT: br i1 [[CMP_N]], label [[DONE:%.*]], label [[SCALAR_PH]] 92; SSE: scalar.ph: 93; SSE-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ 32, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 94; SSE-NEXT: [[BC_MERGE_RDX:%.*]] = phi double [ 0.000000e+00, [[ENTRY]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ] 95; SSE-NEXT: br label [[LOOP:%.*]] 96; SSE: loop: 97; SSE-NEXT: [[I:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[NEXT_ITER:%.*]] ] 98; SSE-NEXT: [[TOT:%.*]] = phi double [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[TOT_NEXT:%.*]], [[NEXT_ITER]] ] 99; SSE-NEXT: [[ADDR:%.*]] = getelementptr double, double* [[ARR]], i32 [[I]] 100; SSE-NEXT: [[NEXTVAL:%.*]] = load double, double* [[ADDR]], align 8 101; SSE-NEXT: [[TST:%.*]] = fcmp fast une double [[NEXTVAL]], 4.200000e+01 102; SSE-NEXT: br i1 [[TST]], label [[DO_ADD:%.*]], label [[NO_ADD:%.*]] 103; SSE: do.add: 104; SSE-NEXT: [[TOT_NEW:%.*]] = fadd fast double [[TOT]], [[NEXTVAL]] 105; SSE-NEXT: br label [[NEXT_ITER]] 106; SSE: no.add: 107; SSE-NEXT: br label [[NEXT_ITER]] 108; SSE: next.iter: 109; SSE-NEXT: [[TOT_NEXT]] = phi double [ [[TOT]], [[NO_ADD]] ], [ [[TOT_NEW]], [[DO_ADD]] ] 110; SSE-NEXT: [[I_NEXT]] = add i32 [[I]], 1 111; SSE-NEXT: [[AGAIN:%.*]] = icmp ult i32 [[I_NEXT]], 32 112; SSE-NEXT: br i1 [[AGAIN]], label [[LOOP]], label [[DONE]], !llvm.loop [[LOOP2:![0-9]+]] 113; SSE: done: 114; SSE-NEXT: [[TOT_NEXT_LCSSA:%.*]] = phi double [ [[TOT_NEXT]], [[NEXT_ITER]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ] 115; SSE-NEXT: ret double [[TOT_NEXT_LCSSA]] 116; 117; AVX-LABEL: @sumIfVector( 118; AVX-NEXT: entry: 119; AVX-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 120; AVX: vector.ph: 121; AVX-NEXT: br label [[VECTOR_BODY:%.*]] 122; AVX: vector.body: 123; AVX-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 124; AVX-NEXT: [[VEC_PHI:%.*]] = phi <4 x double> [ zeroinitializer, [[VECTOR_PH]] ], [ [[PREDPHI:%.*]], [[VECTOR_BODY]] ] 125; AVX-NEXT: [[TMP0:%.*]] = add i32 [[INDEX]], 0 126; AVX-NEXT: [[TMP1:%.*]] = getelementptr double, double* [[ARR:%.*]], i32 [[TMP0]] 127; AVX-NEXT: [[TMP2:%.*]] = getelementptr double, double* [[TMP1]], i32 0 128; AVX-NEXT: [[TMP3:%.*]] = bitcast double* [[TMP2]] to <4 x double>* 129; AVX-NEXT: [[WIDE_LOAD:%.*]] = load <4 x double>, <4 x double>* [[TMP3]], align 8 130; AVX-NEXT: [[TMP4:%.*]] = fcmp fast une <4 x double> [[WIDE_LOAD]], <double 4.200000e+01, double 4.200000e+01, double 4.200000e+01, double 4.200000e+01> 131; AVX-NEXT: [[TMP5:%.*]] = fadd fast <4 x double> [[VEC_PHI]], [[WIDE_LOAD]] 132; AVX-NEXT: [[TMP6:%.*]] = xor <4 x i1> [[TMP4]], <i1 true, i1 true, i1 true, i1 true> 133; AVX-NEXT: [[PREDPHI]] = select <4 x i1> [[TMP4]], <4 x double> [[TMP5]], <4 x double> [[VEC_PHI]] 134; AVX-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4 135; AVX-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], 32 136; AVX-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 137; AVX: middle.block: 138; AVX-NEXT: [[TMP8:%.*]] = call fast double @llvm.vector.reduce.fadd.v4f64(double -0.000000e+00, <4 x double> [[PREDPHI]]) 139; AVX-NEXT: [[CMP_N:%.*]] = icmp eq i32 32, 32 140; AVX-NEXT: br i1 [[CMP_N]], label [[DONE:%.*]], label [[SCALAR_PH]] 141; AVX: scalar.ph: 142; AVX-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ 32, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 143; AVX-NEXT: [[BC_MERGE_RDX:%.*]] = phi double [ 0.000000e+00, [[ENTRY]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ] 144; AVX-NEXT: br label [[LOOP:%.*]] 145; AVX: loop: 146; AVX-NEXT: [[I:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[NEXT_ITER:%.*]] ] 147; AVX-NEXT: [[TOT:%.*]] = phi double [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[TOT_NEXT:%.*]], [[NEXT_ITER]] ] 148; AVX-NEXT: [[ADDR:%.*]] = getelementptr double, double* [[ARR]], i32 [[I]] 149; AVX-NEXT: [[NEXTVAL:%.*]] = load double, double* [[ADDR]], align 8 150; AVX-NEXT: [[TST:%.*]] = fcmp fast une double [[NEXTVAL]], 4.200000e+01 151; AVX-NEXT: br i1 [[TST]], label [[DO_ADD:%.*]], label [[NO_ADD:%.*]] 152; AVX: do.add: 153; AVX-NEXT: [[TOT_NEW:%.*]] = fadd fast double [[TOT]], [[NEXTVAL]] 154; AVX-NEXT: br label [[NEXT_ITER]] 155; AVX: no.add: 156; AVX-NEXT: br label [[NEXT_ITER]] 157; AVX: next.iter: 158; AVX-NEXT: [[TOT_NEXT]] = phi double [ [[TOT]], [[NO_ADD]] ], [ [[TOT_NEW]], [[DO_ADD]] ] 159; AVX-NEXT: [[I_NEXT]] = add i32 [[I]], 1 160; AVX-NEXT: [[AGAIN:%.*]] = icmp ult i32 [[I_NEXT]], 32 161; AVX-NEXT: br i1 [[AGAIN]], label [[LOOP]], label [[DONE]], !llvm.loop [[LOOP2:![0-9]+]] 162; AVX: done: 163; AVX-NEXT: [[TOT_NEXT_LCSSA:%.*]] = phi double [ [[TOT_NEXT]], [[NEXT_ITER]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ] 164; AVX-NEXT: ret double [[TOT_NEXT_LCSSA]] 165; 166entry: 167 br label %loop 168 169loop: 170 %i = phi i32 [0, %entry], [%i.next, %next.iter] 171 %tot = phi double [0.0, %entry], [%tot.next, %next.iter] 172 173 %addr = getelementptr double, double* %arr, i32 %i 174 %nextval = load double, double* %addr 175 176 %tst = fcmp fast une double %nextval, 42.0 177 br i1 %tst, label %do.add, label %no.add 178 179do.add: 180 %tot.new = fadd fast double %tot, %nextval 181 br label %next.iter 182 183no.add: 184 br label %next.iter 185 186next.iter: 187 %tot.next = phi double [%tot, %no.add], [%tot.new, %do.add] 188 %i.next = add i32 %i, 1 189 %again = icmp ult i32 %i.next, 32 190 br i1 %again, label %loop, label %done 191 192done: 193 ret double %tot.next 194} 195 196