Home | History | Annotate | Download | only in OpenMP
      1 // RUN: %clang_cc1 -verify -fopenmp -fnoopenmp-use-tls -DBODY -triple x86_64-unknown-unknown -x c++ -emit-llvm %s -fexceptions -fcxx-exceptions -o - | FileCheck %s
      2 // RUN: %clang_cc1 -fopenmp -fnoopenmp-use-tls -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s
      3 // RUN: %clang_cc1 -fopenmp -fnoopenmp-use-tls -DBODY -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix=CHECK-DEBUG %s
      4 
      5 // RUN: %clang_cc1 -verify -fopenmp -DBODY -triple x86_64-unknown-unknown -x c++ -emit-llvm %s -fexceptions -fcxx-exceptions -o - | FileCheck %s --check-prefix=CHECK-TLS
      6 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s
      7 // RUN: %clang_cc1 -fopenmp -DBODY -x c++ -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -debug-info-kind=limited -std=c++11 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix=CHECK-TLS %s
      8 
      9 // expected-no-diagnostics
     10 // REQUIRES: x86-registered-target
     11 #ifndef HEADER
     12 #define HEADER
     13 // CHECK-DAG: [[IDENT:%.+]] = type { i32, i32, i32, i32, i8* }
     14 // CHECK-DAG: [[S1:%.+]] = type { [[INT:i[0-9]+]] }
     15 // CHECK-DAG: [[S2:%.+]] = type { [[INT]], double }
     16 // CHECK-DAG: [[S3:%.+]] = type { [[INT]], float }
     17 // CHECK-DAG: [[S4:%.+]] = type { [[INT]], [[INT]] }
     18 // CHECK-DAG: [[S5:%.+]] = type { [[INT]], [[INT]], [[INT]] }
     19 // CHECK-DAG: [[SMAIN:%.+]] = type { [[INT]], double, double }
     20 // CHECK-DEBUG-DAG: [[IDENT:%.+]] = type { i32, i32, i32, i32, i8* }
     21 // CHECK-DEBUG-DAG: [[S1:%.+]] = type { [[INT:i[0-9]+]] }
     22 // CHECK-DEBUG-DAG: [[S2:%.+]] = type { [[INT]], double }
     23 // CHECK-DEBUG-DAG: [[S3:%.+]] = type { [[INT]], float }
     24 // CHECK-DEBUG-DAG: [[S4:%.+]] = type { [[INT]], [[INT]] }
     25 // CHECK-DEBUG-DAG: [[S5:%.+]] = type { [[INT]], [[INT]], [[INT]] }
     26 // CHECK-DEBUG-DAG: [[SMAIN:%.+]] = type { [[INT]], double, double }
     27 // CHECK-TLS-DAG: [[S1:%.+]] = type { [[INT:i[0-9]+]] }
     28 // CHECK-TLS-DAG: [[S2:%.+]] = type { [[INT]], double }
     29 // CHECK-TLS-DAG: [[S3:%.+]] = type { [[INT]], float }
     30 // CHECK-TLS-DAG: [[S4:%.+]] = type { [[INT]], [[INT]] }
     31 // CHECK-TLS-DAG: [[S5:%.+]] = type { [[INT]], [[INT]], [[INT]] }
     32 // CHECK-TLS-DAG: [[SMAIN:%.+]] = type { [[INT]], double, double }
     33 
     34 struct S1 {
     35   int a;
     36   S1()
     37       : a(0) {
     38   }
     39   S1(int a)
     40       : a(a) {
     41   }
     42   S1(const S1 &s) {
     43     a = 12 + s.a;
     44   }
     45   ~S1() {
     46     a = 0;
     47   }
     48 };
     49 
     50 struct S2 {
     51   int a;
     52   double b;
     53   S2()
     54       : a(0) {
     55   }
     56   S2(int a)
     57       : a(a) {
     58   }
     59   S2(const S2 &s) {
     60     a = 12 + s.a;
     61   }
     62   ~S2() {
     63     a = 0;
     64   }
     65 };
     66 
     67 struct S3 {
     68   int a;
     69   float b;
     70   S3()
     71       : a(0) {
     72   }
     73   S3(int a)
     74       : a(a) {
     75   }
     76   S3(const S3 &s) {
     77     a = 12 + s.a;
     78   }
     79   ~S3() {
     80     a = 0;
     81   }
     82 };
     83 
     84 struct S4 {
     85   int a, b;
     86   S4()
     87       : a(0) {
     88   }
     89   S4(int a)
     90       : a(a) {
     91   }
     92   S4(const S4 &s) {
     93     a = 12 + s.a;
     94   }
     95   ~S4() {
     96     a = 0;
     97   }
     98 };
     99 
    100 struct S5 {
    101   int a, b, c;
    102   S5()
    103       : a(0) {
    104   }
    105   S5(int a)
    106       : a(a) {
    107   }
    108   S5(const S5 &s) {
    109     a = 12 + s.a;
    110   }
    111   ~S5() {
    112     a = 0;
    113   }
    114 };
    115 
    116 // CHECK-DAG:  [[GS1:@.+]] = internal global [[S1]] zeroinitializer
    117 // CHECK-DAG:  [[GS1]].cache. = common global i8** null
    118 // CHECK-DAG:  [[DEFAULT_LOC:@.+]] = private unnamed_addr constant [[IDENT]] { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([{{[0-9]+}} x i8], [{{[0-9]+}} x i8]* {{@.+}}, i32 0, i32 0) }
    119 // CHECK-DAG:  [[GS2:@.+]] = internal global [[S2]] zeroinitializer
    120 // CHECK-DAG:  [[ARR_X:@.+]] = global [2 x [3 x [[S1]]]] zeroinitializer
    121 // CHECK-DAG:  [[ARR_X]].cache. = common global i8** null
    122 // CHECK-DAG:  [[SM:@.+]] = internal global [[SMAIN]] zeroinitializer
    123 // CHECK-DAG:  [[SM]].cache. = common global i8** null
    124 // CHECK-DAG:  [[STATIC_S:@.+]] = external global [[S3]]
    125 // CHECK-DAG:  [[STATIC_S]].cache. = common global i8** null
    126 // CHECK-DAG:  [[GS3:@.+]] = external global [[S5]]
    127 // CHECK-DAG:  [[GS3]].cache. = common global i8** null
    128 // CHECK-DAG:  [[ST_INT_ST:@.+]] = linkonce_odr global i32 23
    129 // CHECK-DAG:  [[ST_INT_ST]].cache. = common global i8** null
    130 // CHECK-DAG:  [[ST_FLOAT_ST:@.+]] = linkonce_odr global float 2.300000e+01
    131 // CHECK-DAG:  [[ST_FLOAT_ST]].cache. = common global i8** null
    132 // CHECK-DAG:  [[ST_S4_ST:@.+]] = linkonce_odr global %struct.S4 zeroinitializer
    133 // CHECK-DAG:  [[ST_S4_ST]].cache. = common global i8** null
    134 // CHECK-NOT:  .cache. = common global i8** null
    135 // There is no cache for gs2 - it is not threadprivate. Check that there is only
    136 // 8 caches created (for Static::s, gs1, gs3, arr_x, main::sm, ST<int>::st,
    137 // ST<float>::st, ST<S4>::st)
    138 // CHECK-DEBUG-DAG: [[GS1:@.+]] = internal global [[S1]] zeroinitializer
    139 // CHECK-DEBUG-DAG: [[GS2:@.+]] = internal global [[S2]] zeroinitializer
    140 // CHECK-DEBUG-DAG: [[ARR_X:@.+]] = global [2 x [3 x [[S1]]]] zeroinitializer
    141 // CHECK-DEBUG-DAG: [[SM:@.+]] = internal global [[SMAIN]] zeroinitializer
    142 // CHECK-DEBUG-DAG: [[STATIC_S:@.+]] = external global [[S3]]
    143 // CHECK-DEBUG-DAG: [[GS3:@.+]] = external global [[S5]]
    144 // CHECK-DEBUG-DAG: [[ST_INT_ST:@.+]] = linkonce_odr global i32 23
    145 // CHECK-DEBUG-DAG: [[ST_FLOAT_ST:@.+]] = linkonce_odr global float 2.300000e+01
    146 // CHECK-DEBUG-DAG: [[ST_S4_ST:@.+]] = linkonce_odr global %struct.S4 zeroinitializer
    147 // CHECK-DEBUG-DAG: [[LOC1:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;;192;9;;\00"
    148 // CHECK-DEBUG-DAG: [[LOC2:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;;247;9;;\00"
    149 // CHECK-DEBUG-DAG: [[LOC3:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;;334;19;;\00"
    150 // CHECK-DEBUG-DAG: [[LOC4:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;371;9;;\00"
    151 // CHECK-DEBUG-DAG: [[LOC5:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;388;9;;\00"
    152 // CHECK-DEBUG-DAG: [[LOC6:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;410;10;;\00"
    153 // CHECK-DEBUG-DAG: [[LOC7:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;433;10;;\00"
    154 // CHECK-DEBUG-DAG: [[LOC8:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;469;10;;\00"
    155 // CHECK-DEBUG-DAG: [[LOC9:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;498;10;;\00"
    156 // CHECK-DEBUG-DAG: [[LOC10:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;518;10;;\00"
    157 // CHECK-DEBUG-DAG: [[LOC11:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;541;27;;\00"
    158 // CHECK-DEBUG-DAG: [[LOC12:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;main;564;10;;\00"
    159 // CHECK-DEBUG-DAG: [[LOC13:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;684;9;;\00"
    160 // CHECK-DEBUG-DAG: [[LOC14:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;707;10;;\00"
    161 // CHECK-DEBUG-DAG: [[LOC15:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;743;10;;\00"
    162 // CHECK-DEBUG-DAG: [[LOC16:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;772;10;;\00"
    163 // CHECK-DEBUG-DAG: [[LOC17:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;792;10;;\00"
    164 // CHECK-DEBUG-DAG: [[LOC18:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;815;27;;\00"
    165 // CHECK-DEBUG-DAG: [[LOC19:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;foobar;838;10;;\00"
    166 // CHECK-DEBUG-DAG: [[LOC20:@.*]] = private unnamed_addr constant [{{[0-9]+}} x i8] c";{{.*}}threadprivate_codegen.cpp;;306;9;;\00"
    167 // CHECK-TLS-DAG:  [[GS1:@.+]] = internal thread_local global [[S1]] zeroinitializer
    168 // CHECK-TLS-DAG:  [[GS2:@.+]] = internal global [[S2]] zeroinitializer
    169 // CHECK-TLS-DAG:  [[ARR_X:@.+]] = thread_local global [2 x [3 x [[S1]]]] zeroinitializer
    170 // CHECK-TLS-DAG:  [[SM:@.+]] = internal thread_local global [[SMAIN]] zeroinitializer
    171 // CHECK-TLS-DAG:  [[SM_GUARD:@_ZGVZ4mainE2sm]] = internal thread_local global i8 0
    172 // CHECK-TLS-DAG:  [[STATIC_S:@.+]] = external thread_local global [[S3]]
    173 // CHECK-TLS-DAG:  [[GS3:@.+]] = external thread_local global [[S5]]
    174 // CHECK-TLS-DAG:  [[ST_INT_ST:@.+]] = linkonce_odr thread_local global i32 23
    175 // CHECK-TLS-DAG:  [[ST_FLOAT_ST:@.+]] = linkonce_odr thread_local global float 2.300000e+01
    176 // CHECK-TLS-DAG:  [[ST_S4_ST:@.+]] = linkonce_odr thread_local global %struct.S4 zeroinitializer
    177 // CHECK-TLS-DAG:  [[ST_S4_ST_GUARD:@_ZGVN2STI2S4E2stE]] = linkonce_odr thread_local global i64 0
    178 // CHECK-TLS-DAG:  @__tls_guard = internal thread_local global i8 0
    179 // CHECK-TLS-DAG:  @__dso_handle = external global i8
    180 // CHECK-TLS-DAG:  [[GS1_TLS_INIT:@_ZTHL3gs1]] = internal alias void (), void ()* @__tls_init
    181 // CHECK-TLS-DAG:  [[ARR_X_TLS_INIT:@_ZTH5arr_x]] = alias void (), void ()* @__tls_init
    182 // CHECK-TLS-DAG:  [[ST_INT_ST_TLS_INIT:@_ZTHN2STIiE2stE]] = linkonce_odr alias void (), void ()* @__tls_init
    183 // CHECK-TLS-DAG:  [[ST_FLOAT_ST_TLS_INIT:@_ZTHN2STIfE2stE]] = linkonce_odr alias void (), void ()* @__tls_init
    184 // CHECK-TLS-DAG:  [[ST_S4_ST_TLS_INIT:@_ZTHN2STI2S4E2stE]] = linkonce_odr alias void (), void ()* @__tls_init
    185 
    186 struct Static {
    187   static S3 s;
    188 #pragma omp threadprivate(s)
    189 };
    190 
    191 static S1 gs1(5);
    192 #pragma omp threadprivate(gs1)
    193 #pragma omp threadprivate(gs1)
    194 // CHECK:      define {{.*}} [[S1_CTOR:@.*]]([[S1]]* {{.*}},
    195 // CHECK:      define {{.*}} [[S1_DTOR:@.*]]([[S1]]* {{.*}})
    196 // CHECK:      define internal {{.*}}i8* [[GS1_CTOR:@\.__kmpc_global_ctor_\..*]](i8*)
    197 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    198 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    199 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S1]]*
    200 // CHECK-NEXT: call {{.*}} [[S1_CTOR]]([[S1]]* [[RES]], {{.*}} 5)
    201 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    202 // CHECK:      ret i8* [[ARG]]
    203 // CHECK-NEXT: }
    204 // CHECK:      define internal {{.*}}void [[GS1_DTOR:@\.__kmpc_global_dtor_\..*]](i8*)
    205 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    206 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    207 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S1]]*
    208 // CHECK-NEXT: call {{.*}} [[S1_DTOR]]([[S1]]* [[RES]])
    209 // CHECK-NEXT: ret void
    210 // CHECK-NEXT: }
    211 // CHECK:      define internal {{.*}}void [[GS1_INIT:@\.__omp_threadprivate_init_\..*]]()
    212 // CHECK:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[DEFAULT_LOC]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i8* (i8*)* [[GS1_CTOR]], i8* (i8*, i8*)* null, void (i8*)* [[GS1_DTOR]])
    213 // CHECK-NEXT: ret void
    214 // CHECK-NEXT: }
    215 // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    216 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    217 // CHECK-DEBUG:      store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC1]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    218 // CHECK-DEBUG:      @__kmpc_global_thread_num
    219 // CHECK-DEBUG:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[KMPC_LOC_ADDR]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i8* (i8*)* [[GS1_CTOR:@\.__kmpc_global_ctor_\..*]], i8* (i8*, i8*)* null, void (i8*)* [[GS1_DTOR:@\.__kmpc_global_dtor_\..*]])
    220 // CHECK-DEBUG:      define internal {{.*}}i8* [[GS1_CTOR]](i8*)
    221 // CHECK-DEBUG:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    222 // CHECK-DEBUG:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    223 // CHECK-DEBUG:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S1]]*
    224 // CHECK-DEBUG-NEXT: call {{.*}} [[S1_CTOR:@.+]]([[S1]]* [[RES]], {{.*}} 5){{.*}}, !dbg
    225 // CHECK-DEBUG:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    226 // CHECK-DEBUG:      ret i8* [[ARG]]
    227 // CHECK-DEBUG-NEXT: }
    228 // CHECK-DEBUG:      define {{.*}} [[S1_CTOR]]([[S1]]* {{.*}},
    229 // CHECK-DEBUG:      define internal {{.*}}void [[GS1_DTOR]](i8*)
    230 // CHECK-DEBUG:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    231 // CHECK-DEBUG:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    232 // CHECK-DEBUG:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S1]]*
    233 // CHECK-DEBUG-NEXT: call {{.*}} [[S1_DTOR:@.+]]([[S1]]* [[RES]]){{.*}}, !dbg
    234 // CHECK-DEBUG-NEXT: ret void
    235 // CHECK-DEBUG-NEXT: }
    236 // CHECK-DEBUG:      define {{.*}} [[S1_DTOR]]([[S1]]* {{.*}})
    237 static S2 gs2(27);
    238 // CHECK:      define {{.*}} [[S2_CTOR:@.*]]([[S2]]* {{.*}},
    239 // CHECK:      define {{.*}} [[S2_DTOR:@.*]]([[S2]]* {{.*}})
    240 // No another call for S2 constructor because it is not threadprivate
    241 // CHECK-NOT:  call {{.*}} [[S2_CTOR]]([[S2]]*
    242 // CHECK-DEBUG:      define {{.*}} [[S2_CTOR:@.*]]([[S2]]* {{.*}},
    243 // CHECK-DEBUG:      define {{.*}} [[S2_DTOR:@.*]]([[S2]]* {{.*}})
    244 // No another call for S2 constructor because it is not threadprivate
    245 // CHECK-DEBUG-NOT:  call {{.*}} [[S2_CTOR]]([[S2]]*
    246 S1 arr_x[2][3] = { { 1, 2, 3 }, { 4, 5, 6 } };
    247 #pragma omp threadprivate(arr_x)
    248 // CHECK:      define internal {{.*}}i8* [[ARR_X_CTOR:@\.__kmpc_global_ctor_\..*]](i8*)
    249 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    250 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    251 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [2 x [3 x [[S1]]]]*
    252 // CHECK:      [[ARR1:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[RES]], i{{.*}} 0, i{{.*}} 0
    253 // CHECK:      [[ARR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR1]], i{{.*}} 0, i{{.*}} 0
    254 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR]], [[INT]] {{.*}}1)
    255 // CHECK:      [[ARR_ELEMENT:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR]], i{{.*}} 1
    256 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR_ELEMENT]], [[INT]] {{.*}}2)
    257 // CHECK:      [[ARR_ELEMENT2:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_ELEMENT]], i{{.*}} 1
    258 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR_ELEMENT2]], [[INT]] {{.*}}3)
    259 // CHECK:      [[ARR_ELEMENT3:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR1]], i{{.*}} 1
    260 // CHECK:      [[ARR_:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_ELEMENT3]], i{{.*}} 0, i{{.*}} 0
    261 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR_]], [[INT]] {{.*}}4)
    262 // CHECK:      [[ARR_ELEMENT:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_]], i{{.*}} 1
    263 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR_ELEMENT]], [[INT]] {{.*}}5)
    264 // CHECK:      [[ARR_ELEMENT2:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_ELEMENT]], i{{.*}} 1
    265 // CHECK:      invoke {{.*}} [[S1_CTOR]]([[S1]]* [[ARR_ELEMENT2]], [[INT]] {{.*}}6)
    266 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    267 // CHECK:      ret i8* [[ARG]]
    268 // CHECK:      }
    269 // CHECK:      define internal {{.*}}void [[ARR_X_DTOR:@\.__kmpc_global_dtor_\..*]](i8*)
    270 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    271 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    272 // CHECK:      [[ARR_BEGIN:%.*]] = bitcast i8* [[ARG]] to [[S1]]*
    273 // CHECK-NEXT: [[ARR_CUR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_BEGIN]], i{{.*}} 6
    274 // CHECK-NEXT: br label %[[ARR_LOOP:.*]]
    275 // CHECK:      {{.*}}[[ARR_LOOP]]{{.*}}
    276 // CHECK-NEXT: [[ARR_ELEMENTPAST:%.*]] = phi [[S1]]* [ [[ARR_CUR]], {{.*}} ], [ [[ARR_ELEMENT:%.*]], {{.*}} ]
    277 // CHECK-NEXT: [[ARR_ELEMENT:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_ELEMENTPAST]], i{{.*}} -1
    278 // CHECK-NEXT: invoke {{.*}} [[S1_DTOR]]([[S1]]* [[ARR_ELEMENT]])
    279 // CHECK:      [[ARR_DONE:%.*]] = icmp eq [[S1]]* [[ARR_ELEMENT]], [[ARR_BEGIN]]
    280 // CHECK-NEXT: br i1 [[ARR_DONE]], label %[[ARR_EXIT:.*]], label %[[ARR_LOOP]]
    281 // CHECK:      {{.*}}[[ARR_EXIT]]{{.*}}
    282 // CHECK-NEXT: ret void
    283 // CHECK:      }
    284 // CHECK:      define internal {{.*}}void [[ARR_X_INIT:@\.__omp_threadprivate_init_\..*]]()
    285 // CHECK:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[DEFAULT_LOC]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i8* (i8*)* [[ARR_X_CTOR]], i8* (i8*, i8*)* null, void (i8*)* [[ARR_X_DTOR]])
    286 // CHECK-NEXT: ret void
    287 // CHECK-NEXT: }
    288 // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    289 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    290 // CHECK-DEBUG:      store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC2]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    291 // CHECK-DEBUG:      @__kmpc_global_thread_num
    292 // CHECK-DEBUG:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[KMPC_LOC_ADDR]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i8* (i8*)* [[ARR_X_CTOR:@\.__kmpc_global_ctor_\..*]], i8* (i8*, i8*)* null, void (i8*)* [[ARR_X_DTOR:@\.__kmpc_global_dtor_\..*]])
    293 // CHECK-DEBUG:      define internal {{.*}}i8* [[ARR_X_CTOR]](i8*)
    294 // CHECK-DEBUG:      }
    295 // CHECK-DEBUG:      define internal {{.*}}void [[ARR_X_DTOR]](i8*)
    296 // CHECK-DEBUG:      }
    297 extern S5 gs3;
    298 #pragma omp threadprivate(gs3)
    299 // No call for S5 constructor because gs3 has just declaration, not a definition.
    300 // CHECK-NOT:  call {{.*}}([[S5]]*
    301 // CHECK-DEBUG-NOT:  call {{.*}}([[S5]]*
    302 
    303 template <class T>
    304 struct ST {
    305   static T st;
    306 #pragma omp threadprivate(st)
    307 };
    308 
    309 template <class T>
    310 T ST<T>::st(23);
    311 
    312 // CHECK-LABEL:  @main()
    313 // CHECK-DEBUG-LABEL: @main()
    314 int main() {
    315   // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    316   int Res;
    317   struct Smain {
    318     int a;
    319     double b, c;
    320     Smain()
    321         : a(0) {
    322     }
    323     Smain(int a)
    324         : a(a) {
    325     }
    326     Smain(const Smain &s) {
    327       a = 12 + s.a;
    328     }
    329     ~Smain() {
    330       a = 0;
    331     }
    332   };
    333 
    334   static Smain sm(gs1.a);
    335 // CHECK:      [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[DEFAULT_LOC]])
    336 // CHECK:      call {{.*}}i{{.*}} @__cxa_guard_acquire
    337 // CHECK:      call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[DEFAULT_LOC]])
    338 // CHECK:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[DEFAULT_LOC]], i8* bitcast ([[SMAIN]]* [[SM]] to i8*), i8* (i8*)* [[SM_CTOR:@\.__kmpc_global_ctor_\..+]], i8* (i8*, i8*)* null, void (i8*)* [[SM_DTOR:@\.__kmpc_global_dtor_\..+]])
    339 // CHECK:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS1]].cache.)
    340 // CHECK-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    341 // CHECK-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    342 // CHECK-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    343 // CHECK-NEXT: invoke {{.*}} [[SMAIN_CTOR:.*]]([[SMAIN]]* [[SM]], [[INT]] {{.*}}[[GS1_A]])
    344 // CHECK:      call {{.*}}void @__cxa_guard_release
    345 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    346 // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC3]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    347 // CHECK-DEBUG-NEXT: [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[KMPC_LOC_ADDR]])
    348 // CHECK-DEBUG:      call {{.*}}i{{.*}} @__cxa_guard_acquire
    349 // CHECK-DEBUG:      call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[KMPC_LOC_ADDR]])
    350 // CHECK-DEBUG:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[KMPC_LOC_ADDR]], i8* bitcast ([[SMAIN]]* [[SM]] to i8*), i8* (i8*)* [[SM_CTOR:@\.__kmpc_global_ctor_\..+]], i8* (i8*, i8*)* null, void (i8*)* [[SM_DTOR:@\.__kmpc_global_dtor_\..+]])
    351 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    352 // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC3]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    353 // CHECK-DEBUG:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    354 // CHECK-DEBUG-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    355 // CHECK-DEBUG-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    356 // CHECK-DEBUG-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    357 // CHECK-DEBUG-NEXT: invoke {{.*}} [[SMAIN_CTOR:.*]]([[SMAIN]]* [[SM]], [[INT]] {{.*}}[[GS1_A]])
    358 // CHECK-DEBUG:      call {{.*}}void @__cxa_guard_release
    359 // CHECK-TLS:      [[IS_INIT_INT:%.*]] = load i8, i8* [[SM_GUARD]]
    360 // CHECK-TLS-NEXT: [[IS_INIT_BOOL:%.*]] = icmp eq i8 [[IS_INIT_INT]], 0
    361 // CHECK-TLS-NEXT: br i1 [[IS_INIT_BOOL]], label %[[INIT_LABEL:.*]], label %[[INIT_DONE:[^,]+]]{{.*}}
    362 // CHECK-TLS:      [[INIT_LABEL]]
    363 // CHECK-TLS-NEXT: [[GS1_ADDR:%.*]] = call [[S1]]* [[GS1_TLS_INITD:@[^,]+]]
    364 // CHECK-TLS-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i32 0, i32 0
    365 // CHECK-TLS-NEXT: [[GS1_A_VAL:%.*]] = load i32, i32* [[GS1_A_ADDR]]
    366 // CHECK-TLS-NEXT: call void [[SM_CTOR1:@.*]]([[SMAIN]]* [[SM]], i32 [[GS1_A_VAL]])
    367 // CHECK-TLS-NEXT: call i32 @__cxa_thread_atexit(void (i8*)* bitcast (void ([[SMAIN]]*)* [[SM_DTOR1:@.*]] to void (i8*)*), i8* bitcast ([[SMAIN]]* [[SM]] to i8*), i8* @__dso_handle)
    368 // CHECK-TLS-NEXT: store i8 1, i8* [[SM_GUARD]]
    369 // CHECK-TLS-NEXT: br label %[[INIT_DONE]]
    370 // CHECK-TLS:      [[INIT_DONE]]
    371 #pragma omp threadprivate(sm)
    372   // CHECK:      [[STATIC_S_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S3]]* [[STATIC_S]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[STATIC_S]].cache.)
    373   // CHECK-NEXT: [[STATIC_S_ADDR:%.*]] = bitcast i8* [[STATIC_S_TEMP_ADDR]] to [[S3]]*
    374   // CHECK-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    375   // CHECK-NEXT: [[STATIC_S_A:%.*]] = load [[INT]], [[INT]]* [[STATIC_S_A_ADDR]]
    376   // CHECK-NEXT: store [[INT]] [[STATIC_S_A]], [[INT]]* [[RES_ADDR:[^,]+]]
    377   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    378   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC5]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    379   // CHECK-DEBUG-NEXT: [[STATIC_S_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S3]]* [[STATIC_S]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    380   // CHECK-DEBUG-NEXT: [[STATIC_S_ADDR:%.*]] = bitcast i8* [[STATIC_S_TEMP_ADDR]] to [[S3]]*
    381   // CHECK-DEBUG-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    382   // CHECK-DEBUG-NEXT: [[STATIC_S_A:%.*]] = load [[INT]], [[INT]]* [[STATIC_S_A_ADDR]]
    383   // CHECK-DEBUG-NEXT: store [[INT]] [[STATIC_S_A]], [[INT]]* [[RES_ADDR:[^,]+]]
    384   // CHECK-TLS:      [[STATIC_S_ADDR:%.*]] = call [[S3]]* [[STATIC_S_TLS_INITD:@[^,]+]]
    385   // CHECK-TLS-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    386   // CHECK-TLS-NEXT: [[STATIC_S_A:%.*]] = load i32, i32* [[STATIC_S_A_ADDR]]
    387   // CHECK-TLS-NEXT: store i32 [[STATIC_S_A]], i32* [[RES_ADDR:[^,]+]]
    388   Res = Static::s.a;
    389   // CHECK:      [[SM_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[SMAIN]]* [[SM]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[SM]].cache.)
    390   // CHECK-NEXT: [[SM_ADDR:%.*]] = bitcast i8* [[SM_TEMP_ADDR]] to [[SMAIN]]*
    391   // CHECK-NEXT: [[SM_A_ADDR:%.*]] = getelementptr inbounds [[SMAIN]], [[SMAIN]]* [[SM_ADDR]], i{{.*}} 0, i{{.*}} 0
    392   // CHECK-NEXT: [[SM_A:%.*]] = load [[INT]], [[INT]]* [[SM_A_ADDR]]
    393   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    394   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[SM_A]]
    395   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    396   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    397   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC6]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    398   // CHECK-DEBUG-NEXT: [[SM_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[SMAIN]]* [[SM]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    399   // CHECK-DEBUG-NEXT: [[SM_ADDR:%.*]] = bitcast i8* [[SM_TEMP_ADDR]] to [[SMAIN]]*
    400   // CHECK-DEBUG-NEXT: [[SM_A_ADDR:%.*]] = getelementptr inbounds [[SMAIN]], [[SMAIN]]* [[SM_ADDR]], i{{.*}} 0, i{{.*}} 0
    401   // CHECK-DEBUG-NEXT: [[SM_A:%.*]] = load [[INT]], [[INT]]* [[SM_A_ADDR]]
    402   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    403   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[SM_A]]
    404   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    405   // [[SM]] was initialized already, so it can be used directly
    406   // CHECK-TLS:      [[SM_A:%.*]] = load i32, i32* getelementptr inbounds ([[SMAIN]], [[SMAIN]]* [[SM]], i{{.*}} 0, i{{.*}} 0)
    407   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    408   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[SM_A]]
    409   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES_ADDR]]
    410   Res += sm.a;
    411   // CHECK:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS1]].cache.)
    412   // CHECK-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    413   // CHECK-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    414   // CHECK-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    415   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    416   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS1_A]]
    417   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    418   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    419   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC7]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    420   // CHECK-DEBUG-NEXT: [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    421   // CHECK-DEBUG-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    422   // CHECK-DEBUG-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    423   // CHECK-DEBUG-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    424   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    425   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS1_A]]
    426   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    427   // CHECK-TLS:      [[GS1_ADDR:%.*]] = call [[S1]]* [[GS1_TLS_INITD]]
    428   // CHECK-TLS-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    429   // CHECK-TLS-NEXT: [[GS1_A:%.*]] = load i32, i32* [[GS1_A_ADDR]]
    430   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    431   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[GS1_A]]
    432   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES_ADDR]]
    433   Res += gs1.a;
    434   // CHECK:      [[GS2_A:%.*]] = load [[INT]], [[INT]]* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    435   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    436   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS2_A]]
    437   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    438   // CHECK-DEBUG:      [[GS2_A:%.*]] = load [[INT]], [[INT]]* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    439   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    440   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS2_A]]
    441   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    442   // CHECK-TLS:      [[GS2_A:%.*]] = load [[INT]], [[INT]]* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    443   // CHECK-TLS-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    444   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS2_A]]
    445   // CHECK-TLS-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    446   Res += gs2.a;
    447   // CHECK:      [[GS3_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S5]]* [[GS3]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS3]].cache.)
    448   // CHECK-NEXT: [[GS3_ADDR:%.*]] = bitcast i8* [[GS3_TEMP_ADDR]] to [[S5]]*
    449   // CHECK-NEXT: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    450   // CHECK-NEXT: [[GS3_A:%.*]] = load [[INT]], [[INT]]* [[GS3_A_ADDR]]
    451   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    452   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS3_A]]
    453   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    454   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    455   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC8]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    456   // CHECK-DEBUG-NEXT: [[GS3_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S5]]* [[GS3]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    457   // CHECK-DEBUG-NEXT: [[GS3_ADDR:%.*]] = bitcast i8* [[GS3_TEMP_ADDR]] to [[S5]]*
    458   // CHECK-DEBUG-NEXT: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    459   // CHECK-DEBUG-NEXT: [[GS3_A:%.*]] = load [[INT]], [[INT]]* [[GS3_A_ADDR]]
    460   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    461   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS3_A]]
    462   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    463   // CHECK-TLS:      [[GS3_ADDR:%.*]] = call [[S5]]* [[GS3_TLS_INITD:[^,]+]]
    464   // CHECK-TLS-NEXT: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    465   // CHECK-TLS-NEXT: [[GS3_A:%.*]] = load i32, i32* [[GS3_A_ADDR]]
    466   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    467   // CHECK-TLS-NEXT: [[ADD:%.*]] = add nsw i32 [[RES]], [[GS3_A]]
    468   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES_ADDR]]
    469   Res += gs3.a;
    470   // CHECK:      [[ARR_X_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ARR_X]].cache.)
    471   // CHECK-NEXT: [[ARR_X_ADDR:%.*]] = bitcast i8* [[ARR_X_TEMP_ADDR]] to [2 x [3 x [[S1]]]]*
    472   // CHECK-NEXT: [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    473   // CHECK-NEXT: [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    474   // CHECK-NEXT: [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    475   // CHECK-NEXT: [[ARR_X_1_1_A:%.*]] = load [[INT]], [[INT]]* [[ARR_X_1_1_A_ADDR]]
    476   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    477   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ARR_X_1_1_A]]
    478   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    479   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    480   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC9]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    481   // CHECK-DEBUG-NEXT:      [[ARR_X_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    482   // CHECK-DEBUG-NEXT: [[ARR_X_ADDR:%.*]] = bitcast i8* [[ARR_X_TEMP_ADDR]] to [2 x [3 x [[S1]]]]*
    483   // CHECK-DEBUG-NEXT: [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    484   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    485   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    486   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_A:%.*]] = load [[INT]], [[INT]]* [[ARR_X_1_1_A_ADDR]]
    487   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    488   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ARR_X_1_1_A]]
    489   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    490   // CHECK-TLS:       [[ARR_X_ADDR:%.*]] = call [2 x [3 x [[S1]]]]* [[ARR_X_TLS_INITD:[^,]+]]
    491   // CHECK-TLS-NEXT:  [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    492   // CHECK-TLS-NEXT:  [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    493   // CHECK-TLS-NEXT:  [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    494   // CHECK-TLS-NEXT:  [[ARR_X_1_1_A:%.*]] = load i32, i32* [[ARR_X_1_1_A_ADDR]]
    495   // CHECK-TLS-NEXT:  [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    496   // CHECK-TLS-NEXT:  [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[ARR_X_1_1_A]]
    497   // CHECK-TLS-NEXT:  store i32 [[ADD]], i32* [[RES_ADDR]]
    498   Res += arr_x[1][1].a;
    499   // CHECK:      [[ST_INT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[INT]]* [[ST_INT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_INT_ST]].cache.)
    500   // CHECK-NEXT: [[ST_INT_ST_ADDR:%.*]] = bitcast i8* [[ST_INT_ST_TEMP_ADDR]] to [[INT]]*
    501   // CHECK-NEXT: [[ST_INT_ST_VAL:%.*]] = load [[INT]], [[INT]]* [[ST_INT_ST_ADDR]]
    502   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    503   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_INT_ST_VAL]]
    504   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    505   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    506   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC10]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    507   // CHECK-DEBUG-NEXT: [[ST_INT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[INT]]* [[ST_INT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    508   // CHECK-DEBUG-NEXT: [[ST_INT_ST_ADDR:%.*]] = bitcast i8* [[ST_INT_ST_TEMP_ADDR]] to [[INT]]*
    509   // CHECK-DEBUG-NEXT: [[ST_INT_ST_VAL:%.*]] = load [[INT]], [[INT]]* [[ST_INT_ST_ADDR]]
    510   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    511   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_INT_ST_VAL]]
    512   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    513   // CHECK-TLS:       [[ST_INT_ST_ADDR:%.*]] = call i32* [[ST_INT_ST_TLS_INITD:[^,]+]]
    514   // CHECK-TLS-NEXT:  [[ST_INT_ST_VAL:%.*]] = load i32, i32* [[ST_INT_ST_ADDR]]
    515   // CHECK-TLS-NEXT:  [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    516   // CHECK-TLS-NEXT:  [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[ST_INT_ST_VAL]]
    517   // CHECK-TLS-NEXT:  store i32 [[ADD]], i32* [[RES_ADDR]]
    518   Res += ST<int>::st;
    519   // CHECK:      [[ST_FLOAT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast (float* [[ST_FLOAT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_FLOAT_ST]].cache.)
    520   // CHECK-NEXT: [[ST_FLOAT_ST_ADDR:%.*]] = bitcast i8* [[ST_FLOAT_ST_TEMP_ADDR]] to float*
    521   // CHECK-NEXT: [[ST_FLOAT_ST_VAL:%.*]] = load float, float* [[ST_FLOAT_ST_ADDR]]
    522   // CHECK-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]] to [[INT]]
    523   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    524   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[FLOAT_TO_INT_CONV]]
    525   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    526   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    527   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC11]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    528   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast (float* [[ST_FLOAT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    529   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_ADDR:%.*]] = bitcast i8* [[ST_FLOAT_ST_TEMP_ADDR]] to float*
    530   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_VAL:%.*]] = load float, float* [[ST_FLOAT_ST_ADDR]]
    531   // CHECK-DEBUG-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]] to [[INT]]
    532   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    533   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[FLOAT_TO_INT_CONV]]
    534   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    535   // CHECK-TLS: [[ST_FLOAT_ST_ADDR:%.*]]  = call float* [[ST_FLOAT_ST_TLS_INITD:[^,]+]]
    536   // CHECK-TLS-NEXT: [[ST_FLOAT_ST_VAL:%.*]]  = load float, float* [[ST_FLOAT_ST_ADDR]]
    537   // CHECK-TLS-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]]  to i32
    538   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    539   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[FLOAT_TO_INT_CONV]]
    540   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES_ADDR]]
    541   Res += static_cast<int>(ST<float>::st);
    542   // CHECK:      [[ST_S4_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_S4_ST]].cache.)
    543   // CHECK-NEXT: [[ST_S4_ST_ADDR:%.*]] = bitcast i8* [[ST_S4_ST_TEMP_ADDR]] to [[S4]]*
    544   // CHECK-NEXT: [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    545   // CHECK-NEXT: [[ST_S4_ST_A:%.*]] = load [[INT]], [[INT]]* [[ST_S4_ST_A_ADDR]]
    546   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    547   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_S4_ST_A]]
    548   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    549   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    550   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC12]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    551   // CHECK-DEBUG-NEXT: [[ST_S4_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    552   // CHECK-DEBUG-NEXT: [[ST_S4_ST_ADDR:%.*]] = bitcast i8* [[ST_S4_ST_TEMP_ADDR]] to [[S4]]*
    553   // CHECK-DEBUG-NEXT: [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    554   // CHECK-DEBUG-NEXT: [[ST_S4_ST_A:%.*]] = load [[INT]], [[INT]]* [[ST_S4_ST_A_ADDR]]
    555   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    556   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_S4_ST_A]]
    557   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    558   // CHECK-TLS:       [[ST_S4_ST_ADDR:%.*]] = call [[S4]]* [[ST_S4_ST_TLS_INITD:[^,]+]]
    559   // CHECK-TLS-NEXT:  [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    560   // CHECK-TLS-NEXT:  [[ST_S4_ST_A:%.*]] = load i32, i32* [[ST_S4_ST_A_ADDR]]
    561   // CHECK-TLS-NEXT:  [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    562   // CHECK-TLS-NEXT:  [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[ST_S4_ST_A]]
    563   // CHECK-TLS-NEXT:  store i32 [[ADD]], i32* [[RES_ADDR]]
    564   Res += ST<S4>::st.a;
    565   // CHECK:      [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    566   // CHECK-NEXT: ret [[INT]] [[RES]]
    567   // CHECK-DEBUG:      [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    568   // CHECK-DEBUG-NEXT: ret [[INT]] [[RES]]
    569   // CHECK-TLS:      [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    570   // CHECK-TLS-NEXT: ret i32 [[RES]]
    571   return Res;
    572 }
    573 // CHECK: }
    574 
    575 // CHECK:      define internal {{.*}}i8* [[SM_CTOR]](i8*)
    576 // CHECK:      [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[DEFAULT_LOC]])
    577 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    578 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    579 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[SMAIN]]*
    580 // CHECK:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS1]].cache.)
    581 // CHECK-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    582 // CHECK-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    583 // CHECK-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    584 // CHECK-NEXT: call {{.*}} [[SMAIN_CTOR:@.+]]([[SMAIN]]* [[RES]], [[INT]] {{.*}}[[GS1_A]])
    585 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    586 // CHECK-NEXT: ret i8* [[ARG]]
    587 // CHECK-NEXT: }
    588 // CHECK:      define {{.*}} [[SMAIN_CTOR]]([[SMAIN]]* {{.*}},
    589 // CHECK:      define internal {{.*}}void [[SM_DTOR]](i8*)
    590 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    591 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    592 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[SMAIN]]*
    593 // CHECK-NEXT: call {{.*}} [[SMAIN_DTOR:@.+]]([[SMAIN]]* [[RES]])
    594 // CHECK-NEXT: ret void
    595 // CHECK-NEXT: }
    596 // CHECK:      define {{.*}} [[SMAIN_DTOR]]([[SMAIN]]* {{.*}})
    597 // CHECK-DEBUG:      define internal {{.*}}i8* [[SM_CTOR]](i8*)
    598 // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    599 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    600 // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC3]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    601 // CHECK-DEBUG-NEXT: [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[KMPC_LOC_ADDR]])
    602 // CHECK-DEBUG:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    603 // CHECK-DEBUG:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    604 // CHECK-DEBUG:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[SMAIN]]*
    605 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    606 // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC3]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    607 // CHECK-DEBUG:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    608 // CHECK-DEBUG-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    609 // CHECK-DEBUG-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    610 // CHECK-DEBUG-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    611 // CHECK-DEBUG-NEXT: call {{.*}} [[SMAIN_CTOR:@.+]]([[SMAIN]]* [[RES]], [[INT]] {{.*}}[[GS1_A]])
    612 // CHECK-DEBUG:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    613 // CHECK-DEBUG-NEXT: ret i8* [[ARG]]
    614 // CHECK-DEBUG-NEXT: }
    615 // CHECK-DEBUG:      define {{.*}} [[SMAIN_CTOR]]([[SMAIN]]* {{.*}},
    616 // CHECK-DEBUG:      define internal {{.*}} [[SM_DTOR:@.+]](i8*)
    617 // CHECK-DEBUG:      call {{.*}} [[SMAIN_DTOR:@.+]]([[SMAIN]]*
    618 // CHECK-DEBUG:      }
    619 // CHECK-DEBUG:      define {{.*}} [[SMAIN_DTOR]]([[SMAIN]]* {{.*}})
    620 // CHECK-TLS:      define internal [[S1]]* [[GS1_TLS_INITD]] {{#[0-9]+}} {
    621 // CHECK-TLS-NEXT: call void [[GS1_TLS_INIT]]
    622 // CHECK-TLS-NEXT: ret [[S1]]* [[GS1]]
    623 // CHECK-TLS-NEXT: }
    624 // CHECK-TLS: define internal void [[SM_CTOR1]]([[SMAIN]]* %this, i32 {{.*}}) {{.*}} {
    625 // CHECK-TLS: void [[SM_CTOR2:@.*]]([[SMAIN]]* {{.*}}, i32 {{.*}})
    626 // CHECK-TLS: }
    627 // CHECK-TLS: define internal void [[SM_DTOR1]]([[SMAIN]]* %this) {{.*}} {
    628 // CHECK-TLS: void [[SM_DTOR2:@.*]]([[SMAIN]]* {{.*}})
    629 // CHECK-TLS: }
    630 // CHECK-TLS: define {{.*}} [[S3]]* [[STATIC_S_TLS_INITD]]
    631 // CHECK-TLS: call void [[STATIC_S_TLS_INIT:[^,]+]]
    632 // CHECK-TLS: ret [[S3]]* [[STATIC_S]]
    633 // CHECK-TLS: }
    634 // CHECK-TLS: define {{.*}} [[S5]]* [[GS3_TLS_INITD]]
    635 // CHECK-TLS:   call void [[GS3_TLS_INIT:@[^,]+]]
    636 // CHECK-TLS:   ret [[S5]]* [[GS3]]
    637 // CHECK-TLS: }
    638 // CHECK-TLS: define {{.*}} [2 x [3 x [[S1]]]]* [[ARR_X_TLS_INITD]]
    639 // CHECK-TLS:   call void [[ARR_X_TLS_INIT]]
    640 // CHECK-TLS:   ret [2 x [3 x [[S1]]]]* [[ARR_X]]
    641 // CHECK-TLS: }
    642 // CHECK-TLS: define {{.*}} i32* [[ST_INT_ST_TLS_INITD]] {{#[0-9]+}} {
    643 // CHECK-TLS:   call void [[ST_INT_ST_TLS_INIT]]
    644 // CHECK-TLS:   ret i32* [[ST_INT_ST]]
    645 // CHECK-TLS: }
    646 // CHECK-TLS: define {{.*}} float* [[ST_FLOAT_ST_TLS_INITD]] {{#[0-9]+}} {
    647 // CHECK-TLS:   call void [[ST_FLOAT_ST_TLS_INIT]]
    648 // CHECK-TLS:   ret float* [[ST_FLOAT_ST]]
    649 // CHECK-TLS: }
    650 // CHECK-TLS: define {{.*}} [[S4]]* [[ST_S4_ST_TLS_INITD]] {{#[0-9]+}} {
    651 // CHECK-TLS:   call void [[ST_S4_ST_TLS_INIT]]
    652 // CHECK-TLS:   ret [[S4]]* [[ST_S4_ST]]
    653 // CHECK-TLS: }
    654 
    655 #endif
    656 
    657 #ifdef BODY
    658 // CHECK-LABEL:  @{{.*}}foobar{{.*}}()
    659 // CHECK-DEBUG-LABEL: @{{.*}}foobar{{.*}}()
    660 // CHECK-TLS: @{{.*}}foobar{{.*}}()
    661 int foobar() {
    662   // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    663   int Res;
    664   // CHECK:      [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[DEFAULT_LOC]])
    665   // CHECK:      [[STATIC_S_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S3]]* [[STATIC_S]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[STATIC_S]].cache.)
    666   // CHECK-NEXT: [[STATIC_S_ADDR:%.*]] = bitcast i8* [[STATIC_S_TEMP_ADDR]] to [[S3]]*
    667   // CHECK-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    668   // CHECK-NEXT: [[STATIC_S_A:%.*]] = load [[INT]], [[INT]]* [[STATIC_S_A_ADDR]]
    669   // CHECK-NEXT: store [[INT]] [[STATIC_S_A]], [[INT]]* [[RES_ADDR:[^,]+]]
    670   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    671   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC13]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    672   // CHECK-DEBUG-NEXT: [[THREAD_NUM:%.+]] = call {{.*}}i32 @__kmpc_global_thread_num([[IDENT]]* [[KMPC_LOC_ADDR]])
    673   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    674   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC13]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    675   // CHECK-DEBUG-NEXT: [[STATIC_S_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S3]]* [[STATIC_S]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    676   // CHECK-DEBUG-NEXT: [[STATIC_S_ADDR:%.*]] = bitcast i8* [[STATIC_S_TEMP_ADDR]] to [[S3]]*
    677   // CHECK-DEBUG-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    678   // CHECK-DEBUG-NEXT: [[STATIC_S_A:%.*]] = load [[INT]], [[INT]]* [[STATIC_S_A_ADDR]]
    679   // CHECK-DEBUG-NEXT: store [[INT]] [[STATIC_S_A]], [[INT]]* [[RES_ADDR:[^,]+]]
    680   // CHECK-TLS:      [[STATIC_S_ADDR:%.*]]  = call [[S3]]* [[STATIC_S_TLS_INITD]]
    681   // CHECK-TLS-NEXT: [[STATIC_S_A_ADDR:%.*]] = getelementptr inbounds [[S3]], [[S3]]* [[STATIC_S_ADDR]], i{{.*}} 0, i{{.*}} 0
    682   // CHECK-TLS-NEXT: [[STATIC_S_A:%.*]] = load i32, i32* [[STATIC_S_A_ADDR]]
    683   // CHECK-TLS-NEXT: store i32 [[STATIC_S_A]], i32* [[RES_ADDR:[^,]+]]
    684   Res = Static::s.a;
    685   // CHECK:      [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS1]].cache.)
    686   // CHECK-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    687   // CHECK-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    688   // CHECK-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    689   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    690   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS1_A]]
    691   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    692   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    693   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC14]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    694   // CHECK-DEBUG-NEXT: [[GS1_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S1]]* [[GS1]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    695   // CHECK-DEBUG-NEXT: [[GS1_ADDR:%.*]] = bitcast i8* [[GS1_TEMP_ADDR]] to [[S1]]*
    696   // CHECK-DEBUG-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    697   // CHECK-DEBUG-NEXT: [[GS1_A:%.*]] = load [[INT]], [[INT]]* [[GS1_A_ADDR]]
    698   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    699   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS1_A]]
    700   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    701   // CHECK-TLS:      [[GS1_ADDR:%.*]] = call [[S1]]* [[GS1_TLS_INITD]]
    702   // CHECK-TLS-NEXT: [[GS1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[GS1_ADDR]], i{{.*}} 0, i{{.*}} 0
    703   // CHECK-TLS-NEXT: [[GS1_A:%.*]] = load i32, i32* [[GS1_A_ADDR]]
    704   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    705   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[GS1_A]]
    706   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES_ADDR]]
    707   Res += gs1.a;
    708   // CHECK:      [[GS2_A:%.*]] = load [[INT]], [[INT]]* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    709   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    710   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS2_A]]
    711   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    712   // CHECK-DEBUG:      [[GS2_A:%.*]] = load [[INT]], [[INT]]* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    713   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    714   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS2_A]]
    715   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    716   // CHECK-TLS:      [[GS2_A:%.*]] = load i32, i32* getelementptr inbounds ([[S2]], [[S2]]* [[GS2]], i{{.*}} 0, i{{.*}} 0)
    717   // CHECK-TLS-NEXT: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    718   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} i32 [[RES]], [[GS2_A]]
    719   // CHECK-TLS-NEXT: store i32 [[ADD]], i32* [[RES:.+]]
    720   Res += gs2.a;
    721   // CHECK:      [[GS3_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S5]]* [[GS3]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[GS3]].cache.)
    722   // CHECK-NEXT: [[GS3_ADDR:%.*]] = bitcast i8* [[GS3_TEMP_ADDR]] to [[S5]]*
    723   // CHECK-NEXT: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    724   // CHECK-NEXT: [[GS3_A:%.*]] = load [[INT]], [[INT]]* [[GS3_A_ADDR]]
    725   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    726   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS3_A]]
    727   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    728   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    729   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC15]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    730   // CHECK-DEBUG-NEXT: [[GS3_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S5]]* [[GS3]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    731   // CHECK-DEBUG-NEXT: [[GS3_ADDR:%.*]] = bitcast i8* [[GS3_TEMP_ADDR]] to [[S5]]*
    732   // CHECK-DEBUG-NEXT: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    733   // CHECK-DEBUG-NEXT: [[GS3_A:%.*]] = load [[INT]], [[INT]]* [[GS3_A_ADDR]]
    734   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    735   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[GS3_A]]
    736   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    737   // CHECK-TLS:       [[GS3_ADDR:%.*]] = call [[S5]]* [[GS3_TLS_INITD]]
    738   // CHECK-TLS-DEBUG: [[GS3_A_ADDR:%.*]] = getelementptr inbounds [[S5]], [[S5]]* [[GS3_ADDR]], i{{.*}} 0, i{{.*}} 0
    739   // CHECK-TLS-DEBUG: [[GS3_A:%.*]] = load i32, i32* [[GS3_A_ADDR]]
    740   // CHECK-TLS-DEBUG: [[RES:%.*]] = load i32, i32* [[RES_ADDR]]
    741   // CHECK-TLS-DEBUG: [[ADD:%.*]]= add nsw i32 [[RES]], [[GS3_A]]
    742   // CHECK-TLS-DEBUG: store i32 [[ADD]], i32* [[RES_ADDR]]
    743   Res += gs3.a;
    744   // CHECK:      [[ARR_X_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ARR_X]].cache.)
    745   // CHECK-NEXT: [[ARR_X_ADDR:%.*]] = bitcast i8* [[ARR_X_TEMP_ADDR]] to [2 x [3 x [[S1]]]]*
    746   // CHECK-NEXT: [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    747   // CHECK-NEXT: [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    748   // CHECK-NEXT: [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    749   // CHECK-NEXT: [[ARR_X_1_1_A:%.*]] = load [[INT]], [[INT]]* [[ARR_X_1_1_A_ADDR]]
    750   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    751   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ARR_X_1_1_A]]
    752   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    753   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    754   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC16]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    755   // CHECK-DEBUG-NEXT:      [[ARR_X_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([2 x [3 x [[S1]]]]* [[ARR_X]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    756   // CHECK-DEBUG-NEXT: [[ARR_X_ADDR:%.*]] = bitcast i8* [[ARR_X_TEMP_ADDR]] to [2 x [3 x [[S1]]]]*
    757   // CHECK-DEBUG-NEXT: [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    758   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    759   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    760   // CHECK-DEBUG-NEXT: [[ARR_X_1_1_A:%.*]] = load [[INT]], [[INT]]* [[ARR_X_1_1_A_ADDR]]
    761   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    762   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ARR_X_1_1_A]]
    763   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    764   // CHECK-TLS:      [[ARR_X_ADDR:%.*]] = call [2 x [3 x [[S1]]]]* [[ARR_X_TLS_INITD]]
    765   // CHECK-TLS-NEXT: [[ARR_X_1_ADDR:%.*]] = getelementptr inbounds [2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X_ADDR]], i{{.*}} 0, i{{.*}} 1
    766   // CHECK-TLS-NEXT: [[ARR_X_1_1_ADDR:%.*]] = getelementptr inbounds [3 x [[S1]]], [3 x [[S1]]]* [[ARR_X_1_ADDR]], i{{.*}} 0, i{{.*}} 1
    767   // CHECK-TLS-NEXT: [[ARR_X_1_1_A_ADDR:%.*]] = getelementptr inbounds [[S1]], [[S1]]* [[ARR_X_1_1_ADDR]], i{{.*}} 0, i{{.*}} 0
    768   // CHECK-TLS-NEXT: [[ARR_X_1_1_A:%.*]] = load [[INT]], [[INT]]* [[ARR_X_1_1_A_ADDR]]
    769   // CHECK-TLS-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    770   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ARR_X_1_1_A]]
    771   // CHECK-TLS-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    772   Res += arr_x[1][1].a;
    773   // CHECK:      [[ST_INT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[INT]]* [[ST_INT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_INT_ST]].cache.)
    774   // CHECK-NEXT: [[ST_INT_ST_ADDR:%.*]] = bitcast i8* [[ST_INT_ST_TEMP_ADDR]] to [[INT]]*
    775   // CHECK-NEXT: [[ST_INT_ST_VAL:%.*]] = load [[INT]], [[INT]]* [[ST_INT_ST_ADDR]]
    776   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    777   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_INT_ST_VAL]]
    778   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    779   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    780   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC17]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    781   // CHECK-DEBUG-NEXT: [[ST_INT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[INT]]* [[ST_INT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    782   // CHECK-DEBUG-NEXT: [[ST_INT_ST_ADDR:%.*]] = bitcast i8* [[ST_INT_ST_TEMP_ADDR]] to [[INT]]*
    783   // CHECK-DEBUG-NEXT: [[ST_INT_ST_VAL:%.*]] = load [[INT]], [[INT]]* [[ST_INT_ST_ADDR]]
    784   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    785   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_INT_ST_VAL]]
    786   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    787   // CHECK-TLS:      [[ST_INT_ST_ADDR:%.*]] = call i32* [[ST_INT_ST_TLS_INITD]]
    788   // CHECK-TLS-NEXT: [[ST_INT_ST_VAL:%.*]] = load [[INT]], [[INT]]* [[ST_INT_ST_ADDR]]
    789   // CHECK-TLS-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    790   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_INT_ST_VAL]]
    791   // CHECK-TLS-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    792   Res += ST<int>::st;
    793   // CHECK:      [[ST_FLOAT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast (float* [[ST_FLOAT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_FLOAT_ST]].cache.)
    794   // CHECK-NEXT: [[ST_FLOAT_ST_ADDR:%.*]] = bitcast i8* [[ST_FLOAT_ST_TEMP_ADDR]] to float*
    795   // CHECK-NEXT: [[ST_FLOAT_ST_VAL:%.*]] = load float, float* [[ST_FLOAT_ST_ADDR]]
    796   // CHECK-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]] to [[INT]]
    797   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    798   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[FLOAT_TO_INT_CONV]]
    799   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    800   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    801   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC18]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    802   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast (float* [[ST_FLOAT_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    803   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_ADDR:%.*]] = bitcast i8* [[ST_FLOAT_ST_TEMP_ADDR]] to float*
    804   // CHECK-DEBUG-NEXT: [[ST_FLOAT_ST_VAL:%.*]] = load float, float* [[ST_FLOAT_ST_ADDR]]
    805   // CHECK-DEBUG-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]] to [[INT]]
    806   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    807   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[FLOAT_TO_INT_CONV]]
    808   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    809   // CHECK-TLS:      [[ST_FLOAT_ST_ADDR:%.*]] = call float* [[ST_FLOAT_ST_TLS_INITD]]
    810   // CHECK-TLS-NEXT: [[ST_FLOAT_ST_VAL:%.*]] = load float, float* [[ST_FLOAT_ST_ADDR]]
    811   // CHECK-TLS-NEXT: [[FLOAT_TO_INT_CONV:%.*]] = fptosi float [[ST_FLOAT_ST_VAL]] to [[INT]]
    812   // CHECK-TLS-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    813   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[FLOAT_TO_INT_CONV]]
    814   // CHECK-TLS-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    815   Res += static_cast<int>(ST<float>::st);
    816   // CHECK:      [[ST_S4_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[DEFAULT_LOC]], i32 [[THREAD_NUM]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8*** [[ST_S4_ST]].cache.)
    817   // CHECK-NEXT: [[ST_S4_ST_ADDR:%.*]] = bitcast i8* [[ST_S4_ST_TEMP_ADDR]] to [[S4]]*
    818   // CHECK-NEXT: [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    819   // CHECK-NEXT: [[ST_S4_ST_A:%.*]] = load [[INT]], [[INT]]* [[ST_S4_ST_A_ADDR]]
    820   // CHECK-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    821   // CHECK-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_S4_ST_A]]
    822   // CHECK-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    823   // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    824   // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC19]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    825   // CHECK-DEBUG-NEXT: [[ST_S4_ST_TEMP_ADDR:%.*]] = call {{.*}}i8* @__kmpc_threadprivate_cached([[IDENT]]* [[KMPC_LOC_ADDR]], i32 [[THREAD_NUM]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i{{.*}} {{[0-9]+}}, i8***
    826   // CHECK-DEBUG-NEXT: [[ST_S4_ST_ADDR:%.*]] = bitcast i8* [[ST_S4_ST_TEMP_ADDR]] to [[S4]]*
    827   // CHECK-DEBUG-NEXT: [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    828   // CHECK-DEBUG-NEXT: [[ST_S4_ST_A:%.*]] = load [[INT]], [[INT]]* [[ST_S4_ST_A_ADDR]]
    829   // CHECK-DEBUG-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    830   // CHECK-DEBUG-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_S4_ST_A]]
    831   // CHECK-DEBUG-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    832   // CHECK-TLS:      [[ST_S4_ST_ADDR:%.*]] = call [[S4]]* [[ST_S4_ST_TLS_INITD]]
    833   // CHECK-TLS-NEXT: [[ST_S4_ST_A_ADDR:%.*]] = getelementptr inbounds [[S4]], [[S4]]* [[ST_S4_ST_ADDR]], i{{.*}} 0, i{{.*}} 0
    834   // CHECK-TLS-NEXT: [[ST_S4_ST_A:%.*]] = load [[INT]], [[INT]]* [[ST_S4_ST_A_ADDR]]
    835   // CHECK-TLS-NEXT: [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    836   // CHECK-TLS-NEXT: [[ADD:%.*]] = add {{.*}} [[INT]] [[RES]], [[ST_S4_ST_A]]
    837   // CHECK-TLS-NEXT: store [[INT]] [[ADD]], [[INT]]* [[RES:.+]]
    838   Res += ST<S4>::st.a;
    839   // CHECK:      [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    840   // CHECK-NEXT: ret [[INT]] [[RES]]
    841   // CHECK-DEBUG:      [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    842   // CHECK-DEBUG-NEXT: ret [[INT]] [[RES]]
    843   // CHECK-TLS:      [[RES:%.*]] = load [[INT]], [[INT]]* [[RES_ADDR]]
    844   // CHECK-TLS-NEXT: ret [[INT]] [[RES]]
    845   return Res;
    846 }
    847 #endif
    848 
    849 // CHECK:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[DEFAULT_LOC]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i8* (i8*)* [[ST_S4_ST_CTOR:@\.__kmpc_global_ctor_\..+]], i8* (i8*, i8*)* null, void (i8*)* [[ST_S4_ST_DTOR:@\.__kmpc_global_dtor_\..+]])
    850 // CHECK:      define internal {{.*}}i8* [[ST_S4_ST_CTOR]](i8*)
    851 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    852 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    853 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S4]]*
    854 // CHECK-NEXT: call {{.*}} [[S4_CTOR:@.+]]([[S4]]* [[RES]], {{.*}} 23)
    855 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    856 // CHECK-NEXT: ret i8* [[ARG]]
    857 // CHECK-NEXT: }
    858 // CHECK:      define {{.*}} [[S4_CTOR]]([[S4]]* {{.*}},
    859 // CHECK:      define internal {{.*}}void [[ST_S4_ST_DTOR]](i8*)
    860 // CHECK:      store i8* %0, i8** [[ARG_ADDR:%.*]],
    861 // CHECK:      [[ARG:%.+]] = load i8*, i8** [[ARG_ADDR]]
    862 // CHECK:      [[RES:%.*]] = bitcast i8* [[ARG]] to [[S4]]*
    863 // CHECK-NEXT: call {{.*}} [[S4_DTOR:@.+]]([[S4]]* [[RES]])
    864 // CHECK-NEXT: ret void
    865 // CHECK-NEXT: }
    866 // CHECK:      define {{.*}} [[S4_DTOR]]([[S4]]* {{.*}})
    867 // CHECK-DEBUG:      [[KMPC_LOC_ADDR:%.*]] = alloca [[IDENT]]
    868 // CHECK-DEBUG:      [[KMPC_LOC_ADDR_PSOURCE:%.*]] = getelementptr inbounds [[IDENT]], [[IDENT]]* [[KMPC_LOC_ADDR]], i{{.*}} 0, i{{.*}} 4
    869 // CHECK-DEBUG-NEXT: store i8* getelementptr inbounds ([{{.*}} x i8], [{{.*}} x i8]* [[LOC20]], i{{.*}} 0, i{{.*}} 0), i8** [[KMPC_LOC_ADDR_PSOURCE]]
    870 // CHECK-DEBUG:      @__kmpc_global_thread_num
    871 // CHECK-DEBUG:      call {{.*}}void @__kmpc_threadprivate_register([[IDENT]]* [[KMPC_LOC_ADDR]], i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*), i8* (i8*)* [[ST_S4_ST_CTOR:@\.__kmpc_global_ctor_\..+]], i8* (i8*, i8*)* null, void (i8*)* [[ST_S4_ST_DTOR:@\.__kmpc_global_dtor_\..+]])
    872 // CHECK-DEBUG:      define internal {{.*}}i8* [[ST_S4_ST_CTOR]](i8*)
    873 // CHECK-DEBUG:      }
    874 // CHECK-DEBUG:      define {{.*}} [[S4_CTOR:@.*]]([[S4]]* {{.*}},
    875 // CHECK-DEBUG:      define internal {{.*}}void [[ST_S4_ST_DTOR]](i8*)
    876 // CHECK-DEBUG:      }
    877 // CHECK-DEBUG:      define {{.*}} [[S4_DTOR:@.*]]([[S4]]* {{.*}})
    878 
    879 // CHECK:      define internal {{.*}}void {{@.*}}()
    880 // CHECK-DAG:  call {{.*}}void [[GS1_INIT]]()
    881 // CHECK-DAG:  call {{.*}}void [[ARR_X_INIT]]()
    882 // CHECK:      ret void
    883 // CHECK-DEBUG:      define internal {{.*}}void {{@.*}}()
    884 // CHECK-DEBUG:      ret void
    885 
    886 // CHECK-TLS: define internal void [[GS1_CXX_INIT:@.*]]()
    887 // CHECK-TLS: call void [[GS1_CTOR1:@.*]]([[S1]]* [[GS1]], i32 5)
    888 // CHECK-TLS: call i32 @__cxa_thread_atexit(void (i8*)* bitcast (void ([[S1]]*)* [[GS1_DTOR1:.*]] to void (i8*)*), i8* bitcast ([[S1]]* [[GS1]] to i8*)
    889 // CHECK-TLS: }
    890 // CHECK-TLS: define {{.*}}void [[GS1_CTOR1]]([[S1]]* {{.*}}, i32 {{.*}})
    891 // CHECK-TLS: call void [[GS1_CTOR2:@.*]]([[S1]]* {{.*}}, i32 {{.*}})
    892 // CHECK-TLS: }
    893 // CHECK-TLS: define {{.*}}void [[GS1_DTOR1]]([[S1]]* {{.*}})
    894 // CHECK-TLS: call void [[GS1_DTOR2:@.*]]([[S1]]* {{.*}})
    895 // CHECK-TLS: }
    896 // CHECK-TLS: define {{.*}}void [[GS1_CTOR2]]([[S1]]* {{.*}}, i32 {{.*}})
    897 // CHECK-TLS: define {{.*}}void [[GS1_DTOR2]]([[S1]]* {{.*}})
    898 
    899 // CHECK-TLS: define internal void [[GS2_CXX_INIT:@.*]]()
    900 // CHECK-TLS: call void [[GS2_CTOR1:@.*]]([[S2]]* [[GS2]], i32 27)
    901 // CHECK-TLS: call i32 @__cxa_atexit(void (i8*)* bitcast (void ([[S2]]*)* [[GS2_DTOR1:.*]] to void (i8*)*), i8* bitcast ([[S2]]* [[GS2]] to i8*)
    902 // CHECK-TLS: }
    903 // CHECK-TLS: define {{.*}}void [[GS2_CTOR1]]([[S2]]* {{.*}}, i32 {{.*}})
    904 // CHECK-TLS: call void [[GS2_CTOR2:@.*]]([[S2]]* {{.*}}, i32 {{.*}})
    905 // CHECK-TLS: }
    906 // CHECK-TLS: define {{.*}}void [[GS2_DTOR1]]([[S2]]* {{.*}})
    907 // CHECK-TLS: call void [[GS2_DTOR2:@.*]]([[S2]]* {{.*}})
    908 // CHECK-TLS: }
    909 // CHECK-TLS: define {{.*}}void [[GS2_CTOR2]]([[S2]]* {{.*}}, i32 {{.*}})
    910 // CHECK-TLS: define {{.*}}void [[GS2_DTOR2]]([[S2]]* {{.*}})
    911 
    912 // CHECK-TLS: define internal void [[ARR_X_CXX_INIT:@.*]]()
    913 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 0, i{{.*}} 0), i{{.*}} 1)
    914 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 0, i{{.*}} 1), i{{.*}} 2)
    915 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 0, i{{.*}} 2), i{{.*}} 3)
    916 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 1, i{{.*}} 0), i{{.*}} 4)
    917 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 1, i{{.*}} 1), i{{.*}} 5)
    918 // CHECK-TLS: invoke void [[GS1_CTOR1]]([[S1]]* getelementptr inbounds ([2 x [3 x [[S1]]]], [2 x [3 x [[S1]]]]* [[ARR_X]], i{{.*}} 0, i{{.*}} 1, i{{.*}} 2), i{{.*}} 6)
    919 // CHECK-TLS: call i32 @__cxa_thread_atexit(void (i8*)* [[ARR_X_CXX_DTOR:@[^,]+]]
    920 // CHECK-TLS: define internal void [[ARR_X_CXX_DTOR]](i8*)
    921 // CHECK-TLS: void [[GS1_DTOR1]]([[S1]]* {{.*}})
    922 
    923 // CHECK-TLS: define {{.*}}void [[SM_CTOR2]]([[SMAIN]]* {{.*}}, i32 {{.*}})
    924 // CHECK-TLS: define {{.*}}void [[SM_DTOR2]]([[SMAIN]]* {{.*}})
    925 
    926 // CHECK-TLS: define internal void [[ST_S4_ST_CXX_INIT:@.*]]()
    927 // CHECK-TLS: call void [[ST_S4_ST_CTOR1:@.*]]([[S4]]* [[ST_S4_ST]], i32 23)
    928 // CHECK-TLS: call i32 @__cxa_thread_atexit(void (i8*)* bitcast (void ([[S4]]*)* [[ST_S4_ST_DTOR1:.*]] to void (i8*)*), i8* bitcast ([[S4]]* [[ST_S4_ST]] to i8*)
    929 // CHECK-TLS: }
    930 // CHECK-TLS: define {{.*}}void [[ST_S4_ST_CTOR1]]([[S4]]* {{.*}}, i32 {{.*}})
    931 // CHECK-TLS: call void [[ST_S4_ST_CTOR2:@.*]]([[S4]]* {{.*}}, i32 {{.*}})
    932 // CHECK-TLS: }
    933 // CHECK-TLS: define {{.*}}void [[ST_S4_ST_DTOR1]]([[S4]]* {{.*}})
    934 // CHECK-TLS: call void [[ST_S4_ST_DTOR2:@.*]]([[S4]]* {{.*}})
    935 // CHECK-TLS: }
    936 // CHECK-TLS: define {{.*}}void [[ST_S4_ST_CTOR2]]([[S4]]* {{.*}}, i32 {{.*}})
    937 // CHECK-TLS: define {{.*}}void [[ST_S4_ST_DTOR2]]([[S4]]* {{.*}})
    938 
    939 // CHECK-TLS:      define internal void @__tls_init()
    940 // CHECK-TLS:      [[GRD:%.*]] = load i8, i8* @__tls_guard
    941 // CHECK-TLS-NEXT: [[IS_INIT:%.*]] = icmp eq i8 [[GRD]], 0
    942 // CHECK-TLS-NEXT: br i1 [[IS_INIT]], label %[[INIT_LABEL:[^,]+]], label %[[DONE_LABEL:[^,]+]]{{.*}}
    943 // CHECK-TLS:      [[INIT_LABEL]]
    944 // CHECK-TLS-NEXT: store i8 1, i8* @__tls_guard
    945 // CHECK-TLS:      call void [[GS1_CXX_INIT]]
    946 // CHECK-TLS-NOT:  call void [[GS2_CXX_INIT]]
    947 // CHECK-TLS:      call void [[ARR_X_CXX_INIT]]
    948 // CHECK-TLS:      call void [[ST_S4_ST_CXX_INIT]]
    949 // CHECK-TLS:      [[DONE_LABEL]]
    950 
    951 // CHECK-TLS-DAG:      declare {{.*}} void [[GS3_TLS_INIT]]
    952 // CHECK-TLS-DAG:      declare {{.*}} void [[STATIC_S_TLS_INIT]]
    953