1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -slp-vectorizer -S | FileCheck %s 3 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(double *%array1, double *%array2, double *%ptrA, double *%ptrB) { 9; CHECK-LABEL: @splat_loads_double( 10; CHECK-NEXT: entry: 11; CHECK-NEXT: [[GEP_1_0:%.*]] = getelementptr inbounds double, double* [[ARRAY1:%.*]], i64 0 12; CHECK-NEXT: [[GEP_2_0:%.*]] = getelementptr inbounds double, double* [[ARRAY2:%.*]], i64 0 13; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds double, double* [[ARRAY2]], i64 1 14; CHECK-NEXT: [[LD_2_0:%.*]] = load double, double* [[GEP_2_0]], align 8 15; CHECK-NEXT: [[LD_2_1:%.*]] = load double, double* [[GEP_2_1]], align 8 16; CHECK-NEXT: [[TMP0:%.*]] = bitcast double* [[GEP_1_0]] to <2 x double>* 17; CHECK-NEXT: [[TMP1:%.*]] = load <2 x double>, <2 x double>* [[TMP0]], align 8 18; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x double> poison, double [[LD_2_0]], i32 0 19; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x double> [[TMP2]], double [[LD_2_0]], i32 1 20; CHECK-NEXT: [[TMP4:%.*]] = fmul <2 x double> [[TMP1]], [[TMP3]] 21; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x double> poison, double [[LD_2_1]], i32 0 22; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x double> [[TMP5]], double [[LD_2_1]], i32 1 23; CHECK-NEXT: [[TMP7:%.*]] = fmul <2 x double> [[TMP1]], [[TMP6]] 24; CHECK-NEXT: [[TMP8:%.*]] = fadd <2 x double> [[TMP4]], [[TMP7]] 25; CHECK-NEXT: [[TMP9:%.*]] = bitcast double* [[GEP_1_0]] to <2 x double>* 26; CHECK-NEXT: store <2 x double> [[TMP8]], <2 x double>* [[TMP9]], align 8 27; CHECK-NEXT: ret void 28; 29entry: 30 %gep_1_0 = getelementptr inbounds double, double* %array1, i64 0 31 %gep_1_1 = getelementptr inbounds double, double* %array1, i64 1 32 %ld_1_0 = load double, double* %gep_1_0, align 8 33 %ld_1_1 = load double, double* %gep_1_1, align 8 34 35 %gep_2_0 = getelementptr inbounds double, double* %array2, i64 0 36 %gep_2_1 = getelementptr inbounds double, double* %array2, i64 1 37 %ld_2_0 = load double, double* %gep_2_0, align 8 38 %ld_2_1 = load double, double* %gep_2_1, align 8 39 40 %mul0 = fmul double %ld_1_0, %ld_2_0 41 %mul1 = fmul double %ld_1_1, %ld_2_0 42 43 %mul2 = fmul double %ld_1_0, %ld_2_1 44 %mul3 = fmul double %ld_1_1, %ld_2_1 45 46 %add0 = fadd double %mul0, %mul2 47 %add1 = fadd double %mul1, %mul3 48 49 store double %add0, double *%gep_1_0 50 store double %add1, double *%gep_1_1 51 ret void 52} 53 54; Same but with float instead of double 55define void @splat_loads_float(float *%array1, float *%array2, float *%ptrA, float *%ptrB) { 56; CHECK-LABEL: @splat_loads_float( 57; CHECK-NEXT: entry: 58; CHECK-NEXT: [[GEP_1_0:%.*]] = getelementptr inbounds float, float* [[ARRAY1:%.*]], i64 0 59; CHECK-NEXT: [[GEP_2_0:%.*]] = getelementptr inbounds float, float* [[ARRAY2:%.*]], i64 0 60; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds float, float* [[ARRAY2]], i64 1 61; CHECK-NEXT: [[LD_2_0:%.*]] = load float, float* [[GEP_2_0]], align 8 62; CHECK-NEXT: [[LD_2_1:%.*]] = load float, float* [[GEP_2_1]], align 8 63; CHECK-NEXT: [[TMP0:%.*]] = bitcast float* [[GEP_1_0]] to <2 x float>* 64; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, <2 x float>* [[TMP0]], align 8 65; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[LD_2_0]], i32 0 66; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x float> [[TMP2]], float [[LD_2_0]], i32 1 67; CHECK-NEXT: [[TMP4:%.*]] = fmul <2 x float> [[TMP1]], [[TMP3]] 68; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x float> poison, float [[LD_2_1]], i32 0 69; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> [[TMP5]], float [[LD_2_1]], i32 1 70; CHECK-NEXT: [[TMP7:%.*]] = fmul <2 x float> [[TMP1]], [[TMP6]] 71; CHECK-NEXT: [[TMP8:%.*]] = fadd <2 x float> [[TMP4]], [[TMP7]] 72; CHECK-NEXT: [[TMP9:%.*]] = bitcast float* [[GEP_1_0]] to <2 x float>* 73; CHECK-NEXT: store <2 x float> [[TMP8]], <2 x float>* [[TMP9]], align 4 74; CHECK-NEXT: ret void 75; 76entry: 77 %gep_1_0 = getelementptr inbounds float, float* %array1, i64 0 78 %gep_1_1 = getelementptr inbounds float, float* %array1, i64 1 79 %ld_1_0 = load float, float* %gep_1_0, align 8 80 %ld_1_1 = load float, float* %gep_1_1, align 8 81 82 %gep_2_0 = getelementptr inbounds float, float* %array2, i64 0 83 %gep_2_1 = getelementptr inbounds float, float* %array2, i64 1 84 %ld_2_0 = load float, float* %gep_2_0, align 8 85 %ld_2_1 = load float, float* %gep_2_1, align 8 86 87 %mul0 = fmul float %ld_1_0, %ld_2_0 88 %mul1 = fmul float %ld_1_1, %ld_2_0 89 90 %mul2 = fmul float %ld_1_0, %ld_2_1 91 %mul3 = fmul float %ld_1_1, %ld_2_1 92 93 %add0 = fadd float %mul0, %mul2 94 %add1 = fadd float %mul1, %mul3 95 96 store float %add0, float *%gep_1_0 97 store float %add1, float *%gep_1_1 98 ret void 99} 100 101; Same but with i64 102define void @splat_loads_i64(i64 *%array1, i64 *%array2, i64 *%ptrA, i64 *%ptrB) { 103; CHECK-LABEL: @splat_loads_i64( 104; CHECK-NEXT: entry: 105; CHECK-NEXT: [[GEP_1_0:%.*]] = getelementptr inbounds i64, i64* [[ARRAY1:%.*]], i64 0 106; CHECK-NEXT: [[GEP_2_0:%.*]] = getelementptr inbounds i64, i64* [[ARRAY2:%.*]], i64 0 107; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds i64, i64* [[ARRAY2]], i64 1 108; CHECK-NEXT: [[LD_2_0:%.*]] = load i64, i64* [[GEP_2_0]], align 8 109; CHECK-NEXT: [[LD_2_1:%.*]] = load i64, i64* [[GEP_2_1]], align 8 110; CHECK-NEXT: [[TMP0:%.*]] = bitcast i64* [[GEP_1_0]] to <2 x i64>* 111; CHECK-NEXT: [[TMP1:%.*]] = load <2 x i64>, <2 x i64>* [[TMP0]], align 8 112; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x i64> poison, i64 [[LD_2_0]], i32 0 113; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x i64> [[TMP2]], i64 [[LD_2_0]], i32 1 114; CHECK-NEXT: [[TMP4:%.*]] = or <2 x i64> [[TMP1]], [[TMP3]] 115; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> poison, i64 [[LD_2_1]], i32 0 116; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x i64> [[TMP5]], i64 [[LD_2_1]], i32 1 117; CHECK-NEXT: [[TMP7:%.*]] = or <2 x i64> [[TMP1]], [[TMP6]] 118; CHECK-NEXT: [[TMP8:%.*]] = add <2 x i64> [[TMP4]], [[TMP7]] 119; CHECK-NEXT: [[TMP9:%.*]] = bitcast i64* [[GEP_1_0]] to <2 x i64>* 120; CHECK-NEXT: store <2 x i64> [[TMP8]], <2 x i64>* [[TMP9]], align 4 121; CHECK-NEXT: ret void 122; 123entry: 124 %gep_1_0 = getelementptr inbounds i64, i64* %array1, i64 0 125 %gep_1_1 = getelementptr inbounds i64, i64* %array1, i64 1 126 %ld_1_0 = load i64, i64* %gep_1_0, align 8 127 %ld_1_1 = load i64, i64* %gep_1_1, align 8 128 129 %gep_2_0 = getelementptr inbounds i64, i64* %array2, i64 0 130 %gep_2_1 = getelementptr inbounds i64, i64* %array2, i64 1 131 %ld_2_0 = load i64, i64* %gep_2_0, align 8 132 %ld_2_1 = load i64, i64* %gep_2_1, align 8 133 134 %or0 = or i64 %ld_1_0, %ld_2_0 135 %or1 = or i64 %ld_1_1, %ld_2_0 136 137 %or2 = or i64 %ld_1_0, %ld_2_1 138 %or3 = or i64 %ld_1_1, %ld_2_1 139 140 %add0 = add i64 %or0, %or2 141 %add1 = add i64 %or1, %or3 142 143 store i64 %add0, i64 *%gep_1_0 144 store i64 %add1, i64 *%gep_1_1 145 ret void 146} 147 148; Same but with i32 149define void @splat_loads_i32(i32 *%array1, i32 *%array2, i32 *%ptrA, i32 *%ptrB) { 150; CHECK-LABEL: @splat_loads_i32( 151; CHECK-NEXT: entry: 152; CHECK-NEXT: [[GEP_1_0:%.*]] = getelementptr inbounds i32, i32* [[ARRAY1:%.*]], i64 0 153; CHECK-NEXT: [[GEP_2_0:%.*]] = getelementptr inbounds i32, i32* [[ARRAY2:%.*]], i64 0 154; CHECK-NEXT: [[GEP_2_1:%.*]] = getelementptr inbounds i32, i32* [[ARRAY2]], i64 1 155; CHECK-NEXT: [[LD_2_0:%.*]] = load i32, i32* [[GEP_2_0]], align 8 156; CHECK-NEXT: [[LD_2_1:%.*]] = load i32, i32* [[GEP_2_1]], align 8 157; CHECK-NEXT: [[TMP0:%.*]] = bitcast i32* [[GEP_1_0]] to <2 x i32>* 158; CHECK-NEXT: [[TMP1:%.*]] = load <2 x i32>, <2 x i32>* [[TMP0]], align 8 159; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x i32> poison, i32 [[LD_2_0]], i32 0 160; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x i32> [[TMP2]], i32 [[LD_2_0]], i32 1 161; CHECK-NEXT: [[TMP4:%.*]] = or <2 x i32> [[TMP1]], [[TMP3]] 162; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i32> poison, i32 [[LD_2_1]], i32 0 163; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x i32> [[TMP5]], i32 [[LD_2_1]], i32 1 164; CHECK-NEXT: [[TMP7:%.*]] = or <2 x i32> [[TMP1]], [[TMP6]] 165; CHECK-NEXT: [[TMP8:%.*]] = add <2 x i32> [[TMP4]], [[TMP7]] 166; CHECK-NEXT: [[TMP9:%.*]] = bitcast i32* [[GEP_1_0]] to <2 x i32>* 167; CHECK-NEXT: store <2 x i32> [[TMP8]], <2 x i32>* [[TMP9]], align 4 168; CHECK-NEXT: ret void 169; 170entry: 171 %gep_1_0 = getelementptr inbounds i32, i32* %array1, i64 0 172 %gep_1_1 = getelementptr inbounds i32, i32* %array1, i64 1 173 %ld_1_0 = load i32, i32* %gep_1_0, align 8 174 %ld_1_1 = load i32, i32* %gep_1_1, align 8 175 176 %gep_2_0 = getelementptr inbounds i32, i32* %array2, i64 0 177 %gep_2_1 = getelementptr inbounds i32, i32* %array2, i64 1 178 %ld_2_0 = load i32, i32* %gep_2_0, align 8 179 %ld_2_1 = load i32, i32* %gep_2_1, align 8 180 181 %or0 = or i32 %ld_1_0, %ld_2_0 182 %or1 = or i32 %ld_1_1, %ld_2_0 183 184 %or2 = or i32 %ld_1_0, %ld_2_1 185 %or3 = or i32 %ld_1_1, %ld_2_1 186 187 %add0 = add i32 %or0, %or2 188 %add1 = add i32 %or1, %or3 189 190 store i32 %add0, i32 *%gep_1_0 191 store i32 %add1, i32 *%gep_1_1 192 ret void 193} 194