153 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt < %s -passes=loop-vectorize -S | FileCheck %s3 4target datalayout = "e-m:e-i64:64-n32:64"5target triple = "powerpc64le-ibm-linux-gnu"6 7; We expect the loop with double reductions to be interleaved 8 times.8define void @test(ptr %arr, i32 %len) {9; CHECK-LABEL: define void @test(10; CHECK-SAME: ptr [[ARR:%.*]], i32 [[LEN:%.*]]) {11; CHECK-NEXT: [[ENTRY:.*]]:12; CHECK-NEXT: [[CMP4:%.*]] = icmp sgt i32 [[LEN]], 013; CHECK-NEXT: br i1 [[CMP4]], label %[[ITER_CHECK:.*]], label %[[FOR_END:.*]]14; CHECK: [[ITER_CHECK]]:15; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], -116; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[LEN]] to i6417; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP1]], 218; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[VEC_EPILOG_SCALAR_PH:.*]], label %[[VECTOR_MAIN_LOOP_ITER_CHECK:.*]]19; CHECK: [[VECTOR_MAIN_LOOP_ITER_CHECK]]:20; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[TMP1]], 1621; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]]22; CHECK: [[VECTOR_PH]]:23; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP1]], 1624; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP1]], [[N_MOD_VF]]25; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]26; CHECK: [[VECTOR_BODY]]:27; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]28; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP12:%.*]], %[[VECTOR_BODY]] ]29; CHECK-NEXT: [[VEC_PHI2:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP13:%.*]], %[[VECTOR_BODY]] ]30; CHECK-NEXT: [[VEC_PHI3:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP14:%.*]], %[[VECTOR_BODY]] ]31; CHECK-NEXT: [[VEC_PHI4:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP15:%.*]], %[[VECTOR_BODY]] ]32; CHECK-NEXT: [[VEC_PHI5:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP16:%.*]], %[[VECTOR_BODY]] ]33; CHECK-NEXT: [[VEC_PHI6:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP17:%.*]], %[[VECTOR_BODY]] ]34; CHECK-NEXT: [[VEC_PHI7:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP18:%.*]], %[[VECTOR_BODY]] ]35; CHECK-NEXT: [[VEC_PHI8:%.*]] = phi <2 x double> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP19:%.*]], %[[VECTOR_BODY]] ]36; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds double, ptr [[ARR]], i64 [[INDEX]]37; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 238; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 439; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 640; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 841; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 1042; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 1243; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds double, ptr [[TMP3]], i64 1444; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP3]], align 845; CHECK-NEXT: [[WIDE_LOAD9:%.*]] = load <2 x double>, ptr [[TMP5]], align 846; CHECK-NEXT: [[WIDE_LOAD10:%.*]] = load <2 x double>, ptr [[TMP6]], align 847; CHECK-NEXT: [[WIDE_LOAD11:%.*]] = load <2 x double>, ptr [[TMP7]], align 848; CHECK-NEXT: [[WIDE_LOAD12:%.*]] = load <2 x double>, ptr [[TMP8]], align 849; CHECK-NEXT: [[WIDE_LOAD13:%.*]] = load <2 x double>, ptr [[TMP9]], align 850; CHECK-NEXT: [[WIDE_LOAD14:%.*]] = load <2 x double>, ptr [[TMP10]], align 851; CHECK-NEXT: [[WIDE_LOAD15:%.*]] = load <2 x double>, ptr [[TMP11]], align 852; CHECK-NEXT: [[TMP12]] = fadd fast <2 x double> [[WIDE_LOAD]], [[VEC_PHI]]53; CHECK-NEXT: [[TMP13]] = fadd fast <2 x double> [[WIDE_LOAD9]], [[VEC_PHI2]]54; CHECK-NEXT: [[TMP14]] = fadd fast <2 x double> [[WIDE_LOAD10]], [[VEC_PHI3]]55; CHECK-NEXT: [[TMP15]] = fadd fast <2 x double> [[WIDE_LOAD11]], [[VEC_PHI4]]56; CHECK-NEXT: [[TMP16]] = fadd fast <2 x double> [[WIDE_LOAD12]], [[VEC_PHI5]]57; CHECK-NEXT: [[TMP17]] = fadd fast <2 x double> [[WIDE_LOAD13]], [[VEC_PHI6]]58; CHECK-NEXT: [[TMP18]] = fadd fast <2 x double> [[WIDE_LOAD14]], [[VEC_PHI7]]59; CHECK-NEXT: [[TMP19]] = fadd fast <2 x double> [[WIDE_LOAD15]], [[VEC_PHI8]]60; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 1661; CHECK-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]62; CHECK-NEXT: br i1 [[TMP20]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]63; CHECK: [[MIDDLE_BLOCK]]:64; CHECK-NEXT: [[BIN_RDX:%.*]] = fadd fast <2 x double> [[TMP13]], [[TMP12]]65; CHECK-NEXT: [[BIN_RDX16:%.*]] = fadd fast <2 x double> [[TMP14]], [[BIN_RDX]]66; CHECK-NEXT: [[BIN_RDX17:%.*]] = fadd fast <2 x double> [[TMP15]], [[BIN_RDX16]]67; CHECK-NEXT: [[BIN_RDX18:%.*]] = fadd fast <2 x double> [[TMP16]], [[BIN_RDX17]]68; CHECK-NEXT: [[BIN_RDX19:%.*]] = fadd fast <2 x double> [[TMP17]], [[BIN_RDX18]]69; CHECK-NEXT: [[BIN_RDX20:%.*]] = fadd fast <2 x double> [[TMP18]], [[BIN_RDX19]]70; CHECK-NEXT: [[BIN_RDX21:%.*]] = fadd fast <2 x double> [[TMP19]], [[BIN_RDX20]]71; CHECK-NEXT: [[TMP21:%.*]] = call fast double @llvm.vector.reduce.fadd.v2f64(double 0.000000e+00, <2 x double> [[BIN_RDX21]])72; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP1]], [[N_VEC]]73; CHECK-NEXT: br i1 [[CMP_N]], label %[[FOR_END_LOOPEXIT:.*]], label %[[VEC_EPILOG_ITER_CHECK:.*]]74; CHECK: [[VEC_EPILOG_ITER_CHECK]]:75; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_MOD_VF]], 276; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label %[[VEC_EPILOG_SCALAR_PH]], label %[[VEC_EPILOG_PH]], !prof [[PROF3:![0-9]+]]77; CHECK: [[VEC_EPILOG_PH]]:78; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0, %[[VECTOR_MAIN_LOOP_ITER_CHECK]] ]79; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi double [ [[TMP21]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0.000000e+00, %[[VECTOR_MAIN_LOOP_ITER_CHECK]] ]80; CHECK-NEXT: [[N_MOD_VF22:%.*]] = urem i64 [[TMP1]], 281; CHECK-NEXT: [[N_VEC23:%.*]] = sub i64 [[TMP1]], [[N_MOD_VF22]]82; CHECK-NEXT: [[TMP22:%.*]] = insertelement <2 x double> zeroinitializer, double [[BC_MERGE_RDX]], i32 083; CHECK-NEXT: br label %[[VEC_EPILOG_VECTOR_BODY:.*]]84; CHECK: [[VEC_EPILOG_VECTOR_BODY]]:85; CHECK-NEXT: [[TMP23:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], %[[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT27:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]86; CHECK-NEXT: [[VEC_PHI25:%.*]] = phi <2 x double> [ [[TMP22]], %[[VEC_EPILOG_PH]] ], [ [[TMP26:%.*]], %[[VEC_EPILOG_VECTOR_BODY]] ]87; CHECK-NEXT: [[TMP24:%.*]] = getelementptr inbounds double, ptr [[ARR]], i64 [[TMP23]]88; CHECK-NEXT: [[WIDE_LOAD26:%.*]] = load <2 x double>, ptr [[TMP24]], align 889; CHECK-NEXT: [[TMP26]] = fadd fast <2 x double> [[WIDE_LOAD26]], [[VEC_PHI25]]90; CHECK-NEXT: [[INDEX_NEXT27]] = add nuw i64 [[TMP23]], 291; CHECK-NEXT: [[TMP27:%.*]] = icmp eq i64 [[INDEX_NEXT27]], [[N_VEC23]]92; CHECK-NEXT: br i1 [[TMP27]], label %[[VEC_EPILOG_MIDDLE_BLOCK:.*]], label %[[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]93; CHECK: [[VEC_EPILOG_MIDDLE_BLOCK]]:94; CHECK-NEXT: [[TMP28:%.*]] = call fast double @llvm.vector.reduce.fadd.v2f64(double 0.000000e+00, <2 x double> [[TMP26]])95; CHECK-NEXT: [[CMP_N28:%.*]] = icmp eq i64 [[TMP1]], [[N_VEC23]]96; CHECK-NEXT: br i1 [[CMP_N28]], label %[[FOR_END_LOOPEXIT]], label %[[VEC_EPILOG_SCALAR_PH]]97; CHECK: [[VEC_EPILOG_SCALAR_PH]]:98; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC23]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[N_VEC]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0, %[[ITER_CHECK]] ]99; CHECK-NEXT: [[BC_MERGE_RDX29:%.*]] = phi double [ [[TMP28]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[TMP21]], %[[VEC_EPILOG_ITER_CHECK]] ], [ 0.000000e+00, %[[ITER_CHECK]] ]100; CHECK-NEXT: br label %[[FOR_BODY:.*]]101; CHECK: [[FOR_BODY]]:102; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[VEC_EPILOG_SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_BODY]] ]103; CHECK-NEXT: [[REDX_05:%.*]] = phi double [ [[BC_MERGE_RDX29]], %[[VEC_EPILOG_SCALAR_PH]] ], [ [[ADD:%.*]], %[[FOR_BODY]] ]104; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds double, ptr [[ARR]], i64 [[INDVARS_IV]]105; CHECK-NEXT: [[TMP29:%.*]] = load double, ptr [[ARRAYIDX]], align 8106; CHECK-NEXT: [[ADD]] = fadd fast double [[TMP29]], [[REDX_05]]107; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1108; CHECK-NEXT: [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV]] to i32109; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], [[TMP0]]110; CHECK-NEXT: br i1 [[EXITCOND]], label %[[FOR_END_LOOPEXIT]], label %[[FOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]111; CHECK: [[FOR_END_LOOPEXIT]]:112; CHECK-NEXT: [[ADD_LCSSA:%.*]] = phi double [ [[ADD]], %[[FOR_BODY]] ], [ [[TMP21]], %[[MIDDLE_BLOCK]] ], [ [[TMP28]], %[[VEC_EPILOG_MIDDLE_BLOCK]] ]113; CHECK-NEXT: br label %[[FOR_END]]114; CHECK: [[FOR_END]]:115; CHECK-NEXT: [[REDX_0_LCSSA:%.*]] = phi double [ 0.000000e+00, %[[ENTRY]] ], [ [[ADD_LCSSA]], %[[FOR_END_LOOPEXIT]] ]116; CHECK-NEXT: ret void117;118entry:119 %cmp4 = icmp sgt i32 %len, 0120 br i1 %cmp4, label %for.body.lr.ph, label %for.end121 122for.body.lr.ph: ; preds = %entry123 %0 = add i32 %len, -1124 br label %for.body125 126for.body: ; preds = %for.body, %for.body.lr.ph127 %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]128 %redx.05 = phi double [ 0.000000e+00, %for.body.lr.ph ], [ %add, %for.body ]129 %arrayidx = getelementptr inbounds double, ptr %arr, i64 %indvars.iv130 %1 = load double, ptr %arrayidx, align 8131 %add = fadd fast double %1, %redx.05132 %indvars.iv.next = add i64 %indvars.iv, 1133 %lftr.wideiv = trunc i64 %indvars.iv to i32134 %exitcond = icmp eq i32 %lftr.wideiv, %0135 br i1 %exitcond, label %for.end.loopexit, label %for.body136 137for.end.loopexit: ; preds = %for.body138 %add.lcssa = phi double [ %add, %for.body ]139 br label %for.end140 141for.end: ; preds = %for.end.loopexit, %entry142 %redx.0.lcssa = phi double [ 0.000000e+00, %entry ], [ %add.lcssa, %for.end.loopexit ]143 ret void144}145;.146; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}147; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}148; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}149; CHECK: [[PROF3]] = !{!"branch_weights", i32 2, i32 14}150; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}151; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META2]], [[META1]]}152;.153