266 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7-avx | FileCheck %s3 4define void @julia_2xdouble(ptr sret([2 x double]), ptr, ptr, ptr) {5; CHECK-LABEL: @julia_2xdouble(6; CHECK-NEXT: top:7; CHECK-NEXT: [[TMP5:%.*]] = load <2 x double>, ptr [[TMP2:%.*]], align 48; CHECK-NEXT: [[TMP7:%.*]] = load <2 x double>, ptr [[TMP3:%.*]], align 49; CHECK-NEXT: [[TMP8:%.*]] = fmul <2 x double> [[TMP5]], [[TMP7]]10; CHECK-NEXT: [[TMP10:%.*]] = load <2 x double>, ptr [[TMP1:%.*]], align 411; CHECK-NEXT: [[TMP11:%.*]] = fadd <2 x double> [[TMP8]], [[TMP10]]12; CHECK-NEXT: [[TMP12:%.*]] = extractelement <2 x double> [[TMP11]], i32 013; CHECK-NEXT: [[I0:%.*]] = insertvalue [2 x double] undef, double [[TMP12]], 014; CHECK-NEXT: [[TMP13:%.*]] = extractelement <2 x double> [[TMP11]], i32 115; CHECK-NEXT: [[I1:%.*]] = insertvalue [2 x double] [[I0]], double [[TMP13]], 116; CHECK-NEXT: store [2 x double] [[I1]], ptr [[TMP0:%.*]], align 417; CHECK-NEXT: ret void18;19top:20 %x0 = load double, ptr %2, align 421 %y0 = load double, ptr %3, align 422 %m0 = fmul double %x0, %y023 %px1 = getelementptr inbounds [2 x double], ptr %2, i64 0, i64 124 %x1 = load double, ptr %px1, align 425 %py1 = getelementptr inbounds [2 x double], ptr %3, i64 0, i64 126 %y1 = load double, ptr %py1, align 427 %m1 = fmul double %x1, %y128 %z0 = load double, ptr %1, align 429 %a0 = fadd double %m0, %z030 %i0 = insertvalue [2 x double] undef, double %a0, 031 %pz1 = getelementptr inbounds [2 x double], ptr %1, i64 0, i64 132 %z1 = load double, ptr %pz1, align 433 %a1 = fadd double %m1, %z134 %i1 = insertvalue [2 x double] %i0, double %a1, 135 store [2 x double] %i1, ptr %0, align 436 ret void37}38 39define void @julia_4xfloat(ptr sret([4 x float]), ptr, ptr, ptr) {40; CHECK-LABEL: @julia_4xfloat(41; CHECK-NEXT: top:42; CHECK-NEXT: [[TMP5:%.*]] = load <4 x float>, ptr [[TMP2:%.*]], align 443; CHECK-NEXT: [[TMP7:%.*]] = load <4 x float>, ptr [[TMP3:%.*]], align 444; CHECK-NEXT: [[TMP8:%.*]] = fmul <4 x float> [[TMP5]], [[TMP7]]45; CHECK-NEXT: [[TMP10:%.*]] = load <4 x float>, ptr [[TMP1:%.*]], align 446; CHECK-NEXT: [[TMP11:%.*]] = fadd <4 x float> [[TMP8]], [[TMP10]]47; CHECK-NEXT: [[TMP12:%.*]] = extractelement <4 x float> [[TMP11]], i32 048; CHECK-NEXT: [[I0:%.*]] = insertvalue [4 x float] undef, float [[TMP12]], 049; CHECK-NEXT: [[TMP13:%.*]] = extractelement <4 x float> [[TMP11]], i32 150; CHECK-NEXT: [[I1:%.*]] = insertvalue [4 x float] [[I0]], float [[TMP13]], 151; CHECK-NEXT: [[TMP14:%.*]] = extractelement <4 x float> [[TMP11]], i32 252; CHECK-NEXT: [[I2:%.*]] = insertvalue [4 x float] [[I1]], float [[TMP14]], 253; CHECK-NEXT: [[TMP15:%.*]] = extractelement <4 x float> [[TMP11]], i32 354; CHECK-NEXT: [[I3:%.*]] = insertvalue [4 x float] [[I2]], float [[TMP15]], 355; CHECK-NEXT: store [4 x float] [[I3]], ptr [[TMP0:%.*]], align 456; CHECK-NEXT: ret void57;58top:59 %x0 = load float, ptr %2, align 460 %y0 = load float, ptr %3, align 461 %m0 = fmul float %x0, %y062 %px1 = getelementptr inbounds [4 x float], ptr %2, i64 0, i64 163 %x1 = load float, ptr %px1, align 464 %py1 = getelementptr inbounds [4 x float], ptr %3, i64 0, i64 165 %y1 = load float, ptr %py1, align 466 %m1 = fmul float %x1, %y167 %px2 = getelementptr inbounds [4 x float], ptr %2, i64 0, i64 268 %x2 = load float, ptr %px2, align 469 %py2 = getelementptr inbounds [4 x float], ptr %3, i64 0, i64 270 %y2 = load float, ptr %py2, align 471 %m2 = fmul float %x2, %y272 %px3 = getelementptr inbounds [4 x float], ptr %2, i64 0, i64 373 %x3 = load float, ptr %px3, align 474 %py3 = getelementptr inbounds [4 x float], ptr %3, i64 0, i64 375 %y3 = load float, ptr %py3, align 476 %m3 = fmul float %x3, %y377 %z0 = load float, ptr %1, align 478 %a0 = fadd float %m0, %z079 %i0 = insertvalue [4 x float] undef, float %a0, 080 %pz1 = getelementptr inbounds [4 x float], ptr %1, i64 0, i64 181 %z1 = load float, ptr %pz1, align 482 %a1 = fadd float %m1, %z183 %i1 = insertvalue [4 x float] %i0, float %a1, 184 %pz2 = getelementptr inbounds [4 x float], ptr %1, i64 0, i64 285 %z2 = load float, ptr %pz2, align 486 %a2 = fadd float %m2, %z287 %i2 = insertvalue [4 x float] %i1, float %a2, 288 %pz3 = getelementptr inbounds [4 x float], ptr %1, i64 0, i64 389 %z3 = load float, ptr %pz3, align 490 %a3 = fadd float %m3, %z391 %i3 = insertvalue [4 x float] %i2, float %a3, 392 store [4 x float] %i3, ptr %0, align 493 ret void94}95 96define void @julia_load_array_of_float(ptr %a, ptr %b, ptr %c) {97; CHECK-LABEL: @julia_load_array_of_float(98; CHECK-NEXT: top:99; CHECK-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr [[A:%.*]], align 4100; CHECK-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr [[B:%.*]], align 4101; CHECK-NEXT: [[TMP4:%.*]] = fsub <4 x float> [[TMP1]], [[TMP3]]102; CHECK-NEXT: [[TMP5:%.*]] = extractelement <4 x float> [[TMP4]], i32 0103; CHECK-NEXT: [[C_ARR0:%.*]] = insertvalue [4 x float] undef, float [[TMP5]], 0104; CHECK-NEXT: [[TMP6:%.*]] = extractelement <4 x float> [[TMP4]], i32 1105; CHECK-NEXT: [[C_ARR1:%.*]] = insertvalue [4 x float] [[C_ARR0]], float [[TMP6]], 1106; CHECK-NEXT: [[TMP7:%.*]] = extractelement <4 x float> [[TMP4]], i32 2107; CHECK-NEXT: [[C_ARR2:%.*]] = insertvalue [4 x float] [[C_ARR1]], float [[TMP7]], 2108; CHECK-NEXT: [[TMP8:%.*]] = extractelement <4 x float> [[TMP4]], i32 3109; CHECK-NEXT: [[C_ARR3:%.*]] = insertvalue [4 x float] [[C_ARR2]], float [[TMP8]], 3110; CHECK-NEXT: store [4 x float] [[C_ARR3]], ptr [[C:%.*]], align 4111; CHECK-NEXT: ret void112;113top:114 %a_arr = load [4 x float], ptr %a, align 4115 %a0 = extractvalue [4 x float] %a_arr, 0116 %a2 = extractvalue [4 x float] %a_arr, 2117 %a1 = extractvalue [4 x float] %a_arr, 1118 %b_arr = load [4 x float], ptr %b, align 4119 %b0 = extractvalue [4 x float] %b_arr, 0120 %b2 = extractvalue [4 x float] %b_arr, 2121 %b1 = extractvalue [4 x float] %b_arr, 1122 %a3 = extractvalue [4 x float] %a_arr, 3123 %c1 = fsub float %a1, %b1124 %b3 = extractvalue [4 x float] %b_arr, 3125 %c0 = fsub float %a0, %b0126 %c2 = fsub float %a2, %b2127 %c_arr0 = insertvalue [4 x float] undef, float %c0, 0128 %c_arr1 = insertvalue [4 x float] %c_arr0, float %c1, 1129 %c3 = fsub float %a3, %b3130 %c_arr2 = insertvalue [4 x float] %c_arr1, float %c2, 2131 %c_arr3 = insertvalue [4 x float] %c_arr2, float %c3, 3132 store [4 x float] %c_arr3, ptr %c, align 4133 ret void134}135 136define void @julia_load_array_of_i32(ptr %a, ptr %b, ptr %c) {137; CHECK-LABEL: @julia_load_array_of_i32(138; CHECK-NEXT: top:139; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr [[A:%.*]], align 4140; CHECK-NEXT: [[TMP3:%.*]] = load <4 x i32>, ptr [[B:%.*]], align 4141; CHECK-NEXT: [[TMP4:%.*]] = sub <4 x i32> [[TMP1]], [[TMP3]]142; CHECK-NEXT: [[TMP5:%.*]] = extractelement <4 x i32> [[TMP4]], i32 0143; CHECK-NEXT: [[C_ARR0:%.*]] = insertvalue [4 x i32] undef, i32 [[TMP5]], 0144; CHECK-NEXT: [[TMP6:%.*]] = extractelement <4 x i32> [[TMP4]], i32 1145; CHECK-NEXT: [[C_ARR1:%.*]] = insertvalue [4 x i32] [[C_ARR0]], i32 [[TMP6]], 1146; CHECK-NEXT: [[TMP7:%.*]] = extractelement <4 x i32> [[TMP4]], i32 2147; CHECK-NEXT: [[C_ARR2:%.*]] = insertvalue [4 x i32] [[C_ARR1]], i32 [[TMP7]], 2148; CHECK-NEXT: [[TMP8:%.*]] = extractelement <4 x i32> [[TMP4]], i32 3149; CHECK-NEXT: [[C_ARR3:%.*]] = insertvalue [4 x i32] [[C_ARR2]], i32 [[TMP8]], 3150; CHECK-NEXT: store [4 x i32] [[C_ARR3]], ptr [[C:%.*]], align 4151; CHECK-NEXT: ret void152;153top:154 %a_arr = load [4 x i32], ptr %a, align 4155 %a0 = extractvalue [4 x i32] %a_arr, 0156 %a2 = extractvalue [4 x i32] %a_arr, 2157 %a1 = extractvalue [4 x i32] %a_arr, 1158 %b_arr = load [4 x i32], ptr %b, align 4159 %b0 = extractvalue [4 x i32] %b_arr, 0160 %b2 = extractvalue [4 x i32] %b_arr, 2161 %b1 = extractvalue [4 x i32] %b_arr, 1162 %a3 = extractvalue [4 x i32] %a_arr, 3163 %c1 = sub i32 %a1, %b1164 %b3 = extractvalue [4 x i32] %b_arr, 3165 %c0 = sub i32 %a0, %b0166 %c2 = sub i32 %a2, %b2167 %c_arr0 = insertvalue [4 x i32] undef, i32 %c0, 0168 %c_arr1 = insertvalue [4 x i32] %c_arr0, i32 %c1, 1169 %c3 = sub i32 %a3, %b3170 %c_arr2 = insertvalue [4 x i32] %c_arr1, i32 %c2, 2171 %c_arr3 = insertvalue [4 x i32] %c_arr2, i32 %c3, 3172 store [4 x i32] %c_arr3, ptr %c, align 4173 ret void174}175 176; Almost identical to previous test, but for type that should NOT be vectorized.177;178define void @julia_load_array_of_i16(ptr %a, ptr %b, ptr %c) {179; CHECK-LABEL: @julia_load_array_of_i16(180; CHECK-NEXT: top:181; CHECK-NEXT: [[A_ARR:%.*]] = load [4 x i16], ptr [[A:%.*]], align 4182; CHECK-NEXT: [[A0:%.*]] = extractvalue [4 x i16] [[A_ARR]], 0183; CHECK-NEXT: [[A2:%.*]] = extractvalue [4 x i16] [[A_ARR]], 2184; CHECK-NEXT: [[A1:%.*]] = extractvalue [4 x i16] [[A_ARR]], 1185; CHECK-NEXT: [[B_ARR:%.*]] = load [4 x i16], ptr [[B:%.*]], align 4186; CHECK-NEXT: [[B0:%.*]] = extractvalue [4 x i16] [[B_ARR]], 0187; CHECK-NEXT: [[B2:%.*]] = extractvalue [4 x i16] [[B_ARR]], 2188; CHECK-NEXT: [[B1:%.*]] = extractvalue [4 x i16] [[B_ARR]], 1189; CHECK-NEXT: [[A3:%.*]] = extractvalue [4 x i16] [[A_ARR]], 3190; CHECK-NEXT: [[C1:%.*]] = sub i16 [[A1]], [[B1]]191; CHECK-NEXT: [[B3:%.*]] = extractvalue [4 x i16] [[B_ARR]], 3192; CHECK-NEXT: [[C0:%.*]] = sub i16 [[A0]], [[B0]]193; CHECK-NEXT: [[C2:%.*]] = sub i16 [[A2]], [[B2]]194; CHECK-NEXT: [[C_ARR0:%.*]] = insertvalue [4 x i16] undef, i16 [[C0]], 0195; CHECK-NEXT: [[C_ARR1:%.*]] = insertvalue [4 x i16] [[C_ARR0]], i16 [[C1]], 1196; CHECK-NEXT: [[C3:%.*]] = sub i16 [[A3]], [[B3]]197; CHECK-NEXT: [[C_ARR2:%.*]] = insertvalue [4 x i16] [[C_ARR1]], i16 [[C2]], 2198; CHECK-NEXT: [[C_ARR3:%.*]] = insertvalue [4 x i16] [[C_ARR2]], i16 [[C3]], 3199; CHECK-NEXT: store [4 x i16] [[C_ARR3]], ptr [[C:%.*]], align 4200; CHECK-NEXT: ret void201;202top:203 %a_arr = load [4 x i16], ptr %a, align 4204 %a0 = extractvalue [4 x i16] %a_arr, 0205 %a2 = extractvalue [4 x i16] %a_arr, 2206 %a1 = extractvalue [4 x i16] %a_arr, 1207 %b_arr = load [4 x i16], ptr %b, align 4208 %b0 = extractvalue [4 x i16] %b_arr, 0209 %b2 = extractvalue [4 x i16] %b_arr, 2210 %b1 = extractvalue [4 x i16] %b_arr, 1211 %a3 = extractvalue [4 x i16] %a_arr, 3212 %c1 = sub i16 %a1, %b1213 %b3 = extractvalue [4 x i16] %b_arr, 3214 %c0 = sub i16 %a0, %b0215 %c2 = sub i16 %a2, %b2216 %c_arr0 = insertvalue [4 x i16] undef, i16 %c0, 0217 %c_arr1 = insertvalue [4 x i16] %c_arr0, i16 %c1, 1218 %c3 = sub i16 %a3, %b3219 %c_arr2 = insertvalue [4 x i16] %c_arr1, i16 %c2, 2220 %c_arr3 = insertvalue [4 x i16] %c_arr2, i16 %c3, 3221 store [4 x i16] %c_arr3, ptr %c, align 4222 ret void223}224 225%pseudovec = type { float, float, float, float }226 227define void @julia_load_struct_of_float(ptr %a, ptr %b, ptr %c) {228; CHECK-LABEL: @julia_load_struct_of_float(229; CHECK-NEXT: top:230; CHECK-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr [[A:%.*]], align 4231; CHECK-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr [[B:%.*]], align 4232; CHECK-NEXT: [[TMP4:%.*]] = fsub <4 x float> [[TMP1]], [[TMP3]]233; CHECK-NEXT: [[TMP5:%.*]] = extractelement <4 x float> [[TMP4]], i32 0234; CHECK-NEXT: [[C_STRUCT0:%.*]] = insertvalue [[PSEUDOVEC:%.*]] undef, float [[TMP5]], 0235; CHECK-NEXT: [[TMP6:%.*]] = extractelement <4 x float> [[TMP4]], i32 1236; CHECK-NEXT: [[C_STRUCT1:%.*]] = insertvalue [[PSEUDOVEC]] [[C_STRUCT0]], float [[TMP6]], 1237; CHECK-NEXT: [[TMP7:%.*]] = extractelement <4 x float> [[TMP4]], i32 2238; CHECK-NEXT: [[C_STRUCT2:%.*]] = insertvalue [[PSEUDOVEC]] [[C_STRUCT1]], float [[TMP7]], 2239; CHECK-NEXT: [[TMP8:%.*]] = extractelement <4 x float> [[TMP4]], i32 3240; CHECK-NEXT: [[C_STRUCT3:%.*]] = insertvalue [[PSEUDOVEC]] [[C_STRUCT2]], float [[TMP8]], 3241; CHECK-NEXT: store [[PSEUDOVEC]] [[C_STRUCT3]], ptr [[C:%.*]], align 4242; CHECK-NEXT: ret void243;244top:245 %a_struct = load %pseudovec, ptr %a, align 4246 %a0 = extractvalue %pseudovec %a_struct, 0247 %a1 = extractvalue %pseudovec %a_struct, 1248 %b_struct = load %pseudovec, ptr %b, align 4249 %a2 = extractvalue %pseudovec %a_struct, 2250 %b0 = extractvalue %pseudovec %b_struct, 0251 %a3 = extractvalue %pseudovec %a_struct, 3252 %c0 = fsub float %a0, %b0253 %b1 = extractvalue %pseudovec %b_struct, 1254 %b2 = extractvalue %pseudovec %b_struct, 2255 %c1 = fsub float %a1, %b1256 %c_struct0 = insertvalue %pseudovec undef, float %c0, 0257 %b3 = extractvalue %pseudovec %b_struct, 3258 %c3 = fsub float %a3, %b3259 %c_struct1 = insertvalue %pseudovec %c_struct0, float %c1, 1260 %c2 = fsub float %a2, %b2261 %c_struct2 = insertvalue %pseudovec %c_struct1, float %c2, 2262 %c_struct3 = insertvalue %pseudovec %c_struct2, float %c3, 3263 store %pseudovec %c_struct3, ptr %c, align 4264 ret void265}266