170 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S | FileCheck %s3 4target triple = "aarch64--linux-gnu"5 6; This checks that we we prefer splats rather than load vectors + shuffles.7; A load + broadcast can be done efficiently with a single `ld1r` instruction.8define void @splat_loads_double(ptr %array1, ptr %array2, ptr %ptrA, ptr %ptrB) {9; CHECK-LABEL: @splat_loads_double(10; CHECK-NEXT: entry:11; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds double, ptr [[ARRAY2:%.*]], i64 112; CHECK-NEXT: [[LD_2_0:%.*]] = load double, ptr [[ARRAY2]], align 813; CHECK-NEXT: [[LD_2_1:%.*]] = load double, ptr [[GEP_2_1]], align 814; CHECK-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[ARRAY1:%.*]], align 815; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x double> poison, double [[LD_2_0]], i32 016; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x double> [[TMP1]], <2 x double> poison, <2 x i32> zeroinitializer17; CHECK-NEXT: [[TMP3:%.*]] = fmul <2 x double> [[TMP0]], [[TMP2]]18; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x double> poison, double [[LD_2_1]], i32 019; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x double> [[TMP4]], <2 x double> poison, <2 x i32> zeroinitializer20; CHECK-NEXT: [[TMP6:%.*]] = fmul <2 x double> [[TMP0]], [[TMP5]]21; CHECK-NEXT: [[TMP7:%.*]] = fadd <2 x double> [[TMP3]], [[TMP6]]22; CHECK-NEXT: store <2 x double> [[TMP7]], ptr [[ARRAY1]], align 823; CHECK-NEXT: ret void24;25entry:26 %gep_1_1 = getelementptr inbounds double, ptr %array1, i64 127 %ld_1_0 = load double, ptr %array1, align 828 %ld_1_1 = load double, ptr %gep_1_1, align 829 30 %gep_2_1 = getelementptr inbounds double, ptr %array2, i64 131 %ld_2_0 = load double, ptr %array2, align 832 %ld_2_1 = load double, ptr %gep_2_1, align 833 34 %mul0 = fmul double %ld_1_0, %ld_2_035 %mul1 = fmul double %ld_1_1, %ld_2_036 37 %mul2 = fmul double %ld_1_0, %ld_2_138 %mul3 = fmul double %ld_1_1, %ld_2_139 40 %add0 = fadd double %mul0, %mul241 %add1 = fadd double %mul1, %mul342 43 store double %add0, ptr %array144 store double %add1, ptr %gep_1_145 ret void46}47 48; Same but with float instead of double49define void @splat_loads_float(ptr %array1, ptr %array2, ptr %ptrA, ptr %ptrB) {50; CHECK-LABEL: @splat_loads_float(51; CHECK-NEXT: entry:52; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds float, ptr [[ARRAY2:%.*]], i64 153; CHECK-NEXT: [[LD_2_0:%.*]] = load float, ptr [[ARRAY2]], align 854; CHECK-NEXT: [[LD_2_1:%.*]] = load float, ptr [[GEP_2_1]], align 855; CHECK-NEXT: [[TMP0:%.*]] = load <2 x float>, ptr [[ARRAY1:%.*]], align 856; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x float> poison, float [[LD_2_0]], i32 057; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x float> [[TMP1]], <2 x float> poison, <2 x i32> zeroinitializer58; CHECK-NEXT: [[TMP3:%.*]] = fmul <2 x float> [[TMP0]], [[TMP2]]59; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x float> poison, float [[LD_2_1]], i32 060; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> zeroinitializer61; CHECK-NEXT: [[TMP6:%.*]] = fmul <2 x float> [[TMP0]], [[TMP5]]62; CHECK-NEXT: [[TMP7:%.*]] = fadd <2 x float> [[TMP3]], [[TMP6]]63; CHECK-NEXT: store <2 x float> [[TMP7]], ptr [[ARRAY1]], align 464; CHECK-NEXT: ret void65;66entry:67 %gep_1_1 = getelementptr inbounds float, ptr %array1, i64 168 %ld_1_0 = load float, ptr %array1, align 869 %ld_1_1 = load float, ptr %gep_1_1, align 870 71 %gep_2_1 = getelementptr inbounds float, ptr %array2, i64 172 %ld_2_0 = load float, ptr %array2, align 873 %ld_2_1 = load float, ptr %gep_2_1, align 874 75 %mul0 = fmul float %ld_1_0, %ld_2_076 %mul1 = fmul float %ld_1_1, %ld_2_077 78 %mul2 = fmul float %ld_1_0, %ld_2_179 %mul3 = fmul float %ld_1_1, %ld_2_180 81 %add0 = fadd float %mul0, %mul282 %add1 = fadd float %mul1, %mul383 84 store float %add0, ptr %array185 store float %add1, ptr %gep_1_186 ret void87}88 89; Same but with i6490define void @splat_loads_i64(ptr %array1, ptr %array2, ptr %ptrA, ptr %ptrB) {91; CHECK-LABEL: @splat_loads_i64(92; CHECK-NEXT: entry:93; CHECK-NEXT: [[GEP_2_2:%.*]] = getelementptr inbounds i64, ptr [[ARRAY3:%.*]], i64 194; CHECK-NEXT: [[LD_2_2:%.*]] = load i64, ptr [[ARRAY3]], align 895; CHECK-NEXT: [[LD_2_3:%.*]] = load i64, ptr [[GEP_2_2]], align 896; CHECK-NEXT: [[TMP0:%.*]] = load <2 x i64>, ptr [[ARRAY1:%.*]], align 897; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x i64> poison, i64 [[LD_2_2]], i32 098; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x i64> [[TMP1]], <2 x i64> poison, <2 x i32> zeroinitializer99; CHECK-NEXT: [[TMP3:%.*]] = or <2 x i64> [[TMP0]], [[TMP2]]100; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x i64> poison, i64 [[LD_2_3]], i32 0101; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x i64> [[TMP4]], <2 x i64> poison, <2 x i32> zeroinitializer102; CHECK-NEXT: [[TMP6:%.*]] = or <2 x i64> [[TMP0]], [[TMP5]]103; CHECK-NEXT: [[TMP7:%.*]] = add <2 x i64> [[TMP3]], [[TMP6]]104; CHECK-NEXT: store <2 x i64> [[TMP7]], ptr [[ARRAY1]], align 8105; CHECK-NEXT: ret void106;107entry:108 %gep_1_1 = getelementptr inbounds i64, ptr %array1, i64 1109 %ld_1_0 = load i64, ptr %array1, align 8110 %ld_1_1 = load i64, ptr %gep_1_1, align 8111 112 %gep_2_1 = getelementptr inbounds i64, ptr %array2, i64 1113 %ld_2_0 = load i64, ptr %array2, align 8114 %ld_2_1 = load i64, ptr %gep_2_1, align 8115 116 %or0 = or i64 %ld_1_0, %ld_2_0117 %or1 = or i64 %ld_1_1, %ld_2_0118 119 %or2 = or i64 %ld_1_0, %ld_2_1120 %or3 = or i64 %ld_1_1, %ld_2_1121 122 %add0 = add i64 %or0, %or2123 %add1 = add i64 %or1, %or3124 125 store i64 %add0, ptr %array1126 store i64 %add1, ptr %gep_1_1127 ret void128}129 130; Same but with i32131define void @splat_loads_i32(ptr %array1, ptr %array2, ptr %ptrA, ptr %ptrB) {132; CHECK-LABEL: @splat_loads_i32(133; CHECK-NEXT: entry:134; CHECK-NEXT: [[GEP_2_2:%.*]] = getelementptr inbounds i32, ptr [[ARRAY3:%.*]], i64 1135; CHECK-NEXT: [[LD_2_2:%.*]] = load i32, ptr [[ARRAY3]], align 8136; CHECK-NEXT: [[LD_2_3:%.*]] = load i32, ptr [[GEP_2_2]], align 8137; CHECK-NEXT: [[TMP0:%.*]] = load <2 x i32>, ptr [[ARRAY1:%.*]], align 8138; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x i32> poison, i32 [[LD_2_2]], i32 0139; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <2 x i32> [[TMP1]], <2 x i32> poison, <2 x i32> zeroinitializer140; CHECK-NEXT: [[TMP3:%.*]] = or <2 x i32> [[TMP0]], [[TMP2]]141; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x i32> poison, i32 [[LD_2_3]], i32 0142; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x i32> [[TMP4]], <2 x i32> poison, <2 x i32> zeroinitializer143; CHECK-NEXT: [[TMP6:%.*]] = or <2 x i32> [[TMP0]], [[TMP5]]144; CHECK-NEXT: [[TMP7:%.*]] = add <2 x i32> [[TMP3]], [[TMP6]]145; CHECK-NEXT: store <2 x i32> [[TMP7]], ptr [[ARRAY1]], align 4146; CHECK-NEXT: ret void147;148entry:149 %gep_1_1 = getelementptr inbounds i32, ptr %array1, i64 1150 %ld_1_0 = load i32, ptr %array1, align 8151 %ld_1_1 = load i32, ptr %gep_1_1, align 8152 153 %gep_2_1 = getelementptr inbounds i32, ptr %array2, i64 1154 %ld_2_0 = load i32, ptr %array2, align 8155 %ld_2_1 = load i32, ptr %gep_2_1, align 8156 157 %or0 = or i32 %ld_1_0, %ld_2_0158 %or1 = or i32 %ld_1_1, %ld_2_0159 160 %or2 = or i32 %ld_1_0, %ld_2_1161 %or3 = or i32 %ld_1_1, %ld_2_1162 163 %add0 = add i32 %or0, %or2164 %add1 = add i32 %or1, %or3165 166 store i32 %add0, ptr %array1167 store i32 %add1, ptr %gep_1_1168 ret void169}170