1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -passes=loop-vectorize -mtriple=thumbv8-unknown-unknown -mcpu=cortex-a53 -S | FileCheck %s 3 4; This test is reduced from SPECFP 2006 482.sphinx. 5; We expect vectorization with <2 x double> and <2 x float> ops. 6; See https://bugs.llvm.org/show_bug.cgi?id=36280 for more details. 7 8 9target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64" 10 11@a = external global i32 12@v = external global i32 13@mm = external global ptr 14@vv = external global ptr 15@ll = external global ptr 16 17define i32 @test(ptr nocapture readonly %x) { 18; CHECK-LABEL: @test( 19; CHECK-NEXT: entry: 20; CHECK-NEXT: [[T:%.*]] = load i32, ptr @v, align 8 21; CHECK-NEXT: [[T1:%.*]] = load i32, ptr @a, align 4 22; CHECK-NEXT: br label [[OUTERLOOP:%.*]] 23; CHECK: outerloop: 24; CHECK-NEXT: [[T2:%.*]] = phi i32 [ [[V17:%.*]], [[OUTEREND:%.*]] ], [ [[T1]], [[ENTRY:%.*]] ] 25; CHECK-NEXT: [[J_0136:%.*]] = phi i32 [ [[INC144:%.*]], [[OUTEREND]] ], [ 0, [[ENTRY]] ] 26; CHECK-NEXT: [[SCORE_1135:%.*]] = phi i32 [ [[CALL142:%.*]], [[OUTEREND]] ], [ -939524096, [[ENTRY]] ] 27; CHECK-NEXT: [[T3:%.*]] = load ptr, ptr @mm, align 4 28; CHECK-NEXT: [[ARRAYIDX109:%.*]] = getelementptr inbounds ptr, ptr [[T3]], i32 [[T2]] 29; CHECK-NEXT: [[T4:%.*]] = load ptr, ptr [[ARRAYIDX109]], align 4 30; CHECK-NEXT: [[T5:%.*]] = load ptr, ptr @vv, align 4 31; CHECK-NEXT: [[ARRAYIDX111:%.*]] = getelementptr inbounds ptr, ptr [[T5]], i32 [[T2]] 32; CHECK-NEXT: [[T6:%.*]] = load ptr, ptr [[ARRAYIDX111]], align 4 33; CHECK-NEXT: [[T7:%.*]] = load ptr, ptr @ll, align 4 34; CHECK-NEXT: [[ARRAYIDX113:%.*]] = getelementptr inbounds float, ptr [[T7]], i32 [[T2]] 35; CHECK-NEXT: [[T8:%.*]] = load float, ptr [[ARRAYIDX113]], align 4 36; CHECK-NEXT: [[CONV114:%.*]] = fpext float [[T8]] to double 37; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[T]], 2 38; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 39; CHECK: vector.ph: 40; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[T]], 2 41; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[T]], [[N_MOD_VF]] 42; CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x double> zeroinitializer, double [[CONV114]], i32 0 43; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 44; CHECK: vector.body: 45; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 46; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x double> [ [[TMP0]], [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ] 47; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 0 48; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds float, ptr [[X:%.*]], i32 [[TMP1]] 49; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds float, ptr [[TMP2]], i32 0 50; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x float>, ptr [[TMP3]], align 4 51; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds float, ptr [[T4]], i32 [[TMP1]] 52; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds float, ptr [[TMP4]], i32 0 53; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x float>, ptr [[TMP5]], align 4 54; CHECK-NEXT: [[TMP6:%.*]] = fsub fast <2 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]] 55; CHECK-NEXT: [[TMP7:%.*]] = fpext <2 x float> [[TMP6]] to <2 x double> 56; CHECK-NEXT: [[TMP8:%.*]] = fmul fast <2 x double> [[TMP7]], [[TMP7]] 57; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds float, ptr [[T6]], i32 [[TMP1]] 58; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds float, ptr [[TMP9]], i32 0 59; CHECK-NEXT: [[WIDE_LOAD2:%.*]] = load <2 x float>, ptr [[TMP10]], align 4 60; CHECK-NEXT: [[TMP11:%.*]] = fpext <2 x float> [[WIDE_LOAD2]] to <2 x double> 61; CHECK-NEXT: [[TMP12:%.*]] = fmul fast <2 x double> [[TMP8]], [[TMP11]] 62; CHECK-NEXT: [[TMP13]] = fsub fast <2 x double> [[VEC_PHI]], [[TMP12]] 63; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2 64; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 65; CHECK-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 66; CHECK: middle.block: 67; CHECK-NEXT: [[TMP15:%.*]] = call fast double @llvm.vector.reduce.fadd.v2f64(double 0.000000e+00, <2 x double> [[TMP13]]) 68; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[T]], [[N_VEC]] 69; CHECK-NEXT: br i1 [[CMP_N]], label [[OUTEREND]], label [[SCALAR_PH]] 70; CHECK: scalar.ph: 71; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[OUTERLOOP]] ] 72; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi double [ [[TMP15]], [[MIDDLE_BLOCK]] ], [ [[CONV114]], [[OUTERLOOP]] ] 73; CHECK-NEXT: br label [[INNERLOOP:%.*]] 74; CHECK: innerloop: 75; CHECK-NEXT: [[I_2132:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC129:%.*]], [[INNERLOOP]] ] 76; CHECK-NEXT: [[DVAL1_4131:%.*]] = phi double [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[SUB127:%.*]], [[INNERLOOP]] ] 77; CHECK-NEXT: [[ARRAYIDX119:%.*]] = getelementptr inbounds float, ptr [[X]], i32 [[I_2132]] 78; CHECK-NEXT: [[T9:%.*]] = load float, ptr [[ARRAYIDX119]], align 4 79; CHECK-NEXT: [[ARRAYIDX120:%.*]] = getelementptr inbounds float, ptr [[T4]], i32 [[I_2132]] 80; CHECK-NEXT: [[T10:%.*]] = load float, ptr [[ARRAYIDX120]], align 4 81; CHECK-NEXT: [[SUB121:%.*]] = fsub fast float [[T9]], [[T10]] 82; CHECK-NEXT: [[CONV122:%.*]] = fpext float [[SUB121]] to double 83; CHECK-NEXT: [[MUL123:%.*]] = fmul fast double [[CONV122]], [[CONV122]] 84; CHECK-NEXT: [[ARRAYIDX124:%.*]] = getelementptr inbounds float, ptr [[T6]], i32 [[I_2132]] 85; CHECK-NEXT: [[T11:%.*]] = load float, ptr [[ARRAYIDX124]], align 4 86; CHECK-NEXT: [[CONV125:%.*]] = fpext float [[T11]] to double 87; CHECK-NEXT: [[MUL126:%.*]] = fmul fast double [[MUL123]], [[CONV125]] 88; CHECK-NEXT: [[SUB127]] = fsub fast double [[DVAL1_4131]], [[MUL126]] 89; CHECK-NEXT: [[INC129]] = add nuw nsw i32 [[I_2132]], 1 90; CHECK-NEXT: [[EXITCOND143:%.*]] = icmp eq i32 [[INC129]], [[T]] 91; CHECK-NEXT: br i1 [[EXITCOND143]], label [[OUTEREND]], label [[INNERLOOP]], !llvm.loop [[LOOP3:![0-9]+]] 92; CHECK: outerend: 93; CHECK-NEXT: [[SUB127_LCSSA:%.*]] = phi double [ [[SUB127]], [[INNERLOOP]] ], [ [[TMP15]], [[MIDDLE_BLOCK]] ] 94; CHECK-NEXT: [[CONV138:%.*]] = fptosi double [[SUB127_LCSSA]] to i32 95; CHECK-NEXT: [[CALL142]] = add nuw nsw i32 [[SCORE_1135]], [[CONV138]] 96; CHECK-NEXT: [[INC144]] = add nuw nsw i32 [[J_0136]], 1 97; CHECK-NEXT: [[ARRAYIDX102:%.*]] = getelementptr inbounds i32, ptr @a, i32 [[INC144]] 98; CHECK-NEXT: [[V17]] = load i32, ptr [[ARRAYIDX102]], align 4 99; CHECK-NEXT: [[CMP103:%.*]] = icmp sgt i32 [[V17]], -1 100; CHECK-NEXT: br i1 [[CMP103]], label [[OUTERLOOP]], label [[EXIT:%.*]] 101; CHECK: exit: 102; CHECK-NEXT: ret i32 [[CALL142]] 103; 104entry: 105 %t = load i32, ptr @v, align 8 106 %t1 = load i32, ptr @a, align 4 107 br label %outerloop 108 109outerloop: 110 %t2 = phi i32 [ %v17, %outerend ], [ %t1, %entry ] 111 %j.0136 = phi i32 [ %inc144, %outerend ], [ 0, %entry ] 112 %score.1135 = phi i32 [ %call142, %outerend ], [ -939524096, %entry ] 113 %t3 = load ptr, ptr @mm, align 4 114 %arrayidx109 = getelementptr inbounds ptr, ptr %t3, i32 %t2 115 %t4 = load ptr, ptr %arrayidx109, align 4 116 %t5 = load ptr, ptr @vv, align 4 117 %arrayidx111 = getelementptr inbounds ptr, ptr %t5, i32 %t2 118 %t6 = load ptr, ptr %arrayidx111, align 4 119 %t7 = load ptr, ptr @ll, align 4 120 %arrayidx113 = getelementptr inbounds float, ptr %t7, i32 %t2 121 %t8 = load float, ptr %arrayidx113, align 4 122 %conv114 = fpext float %t8 to double 123 br label %innerloop 124 125innerloop: 126 %i.2132 = phi i32 [ 0, %outerloop ], [ %inc129, %innerloop ] 127 %dval1.4131 = phi double [ %conv114, %outerloop ], [ %sub127, %innerloop ] 128 %arrayidx119 = getelementptr inbounds float, ptr %x, i32 %i.2132 129 %t9 = load float, ptr %arrayidx119, align 4 130 %arrayidx120 = getelementptr inbounds float, ptr %t4, i32 %i.2132 131 %t10 = load float, ptr %arrayidx120, align 4 132 %sub121 = fsub fast float %t9, %t10 133 %conv122 = fpext float %sub121 to double 134 %mul123 = fmul fast double %conv122, %conv122 135 %arrayidx124 = getelementptr inbounds float, ptr %t6, i32 %i.2132 136 %t11 = load float, ptr %arrayidx124, align 4 137 %conv125 = fpext float %t11 to double 138 %mul126 = fmul fast double %mul123, %conv125 139 %sub127 = fsub fast double %dval1.4131, %mul126 140 %inc129 = add nuw nsw i32 %i.2132, 1 141 %exitcond143 = icmp eq i32 %inc129, %t 142 br i1 %exitcond143, label %outerend, label %innerloop 143 144outerend: 145 %sub127.lcssa = phi double [ %sub127, %innerloop ] 146 %conv138 = fptosi double %sub127.lcssa to i32 147 %call142 = add nuw nsw i32 %score.1135, %conv138 148 %inc144 = add nuw nsw i32 %j.0136, 1 149 %arrayidx102 = getelementptr inbounds i32, ptr @a, i32 %inc144 150 %v17 = load i32, ptr %arrayidx102, align 4 151 %cmp103 = icmp sgt i32 %v17, -1 152 br i1 %cmp103, label %outerloop, label %exit 153 154exit: 155 ret i32 %call142 156} 157 158