1 // RUN: %clang_cc1 -verify -fopenmp -x c++ -triple x86_64-unknown-unknown -emit-llvm %s -fexceptions -fcxx-exceptions -o - | FileCheck %s 2 // RUN: %clang_cc1 -fopenmp -x c++ -std=c++11 -triple x86_64-unknown-unknown -fexceptions -fcxx-exceptions -emit-pch -o %t %s 3 // RUN: %clang_cc1 -fopenmp -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 %s 4 // RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fexceptions -fcxx-exceptions -debug-info-kind=line-tables-only -x c++ -emit-llvm %s -o - | FileCheck %s --check-prefix=TERM_DEBUG 5 // REQUIRES: x86-registered-target 6 // expected-no-diagnostics 7 #ifndef HEADER 8 #define HEADER 9 10 // CHECK: [[SS_TY:%.+]] = type { i32 } 11 12 long long get_val() { return 0; } 13 double *g_ptr; 14 15 // CHECK-LABEL: define {{.*void}} @{{.*}}simple{{.*}}(float* {{.+}}, float* {{.+}}, float* {{.+}}, float* {{.+}}) 16 void simple(float *a, float *b, float *c, float *d) { 17 #pragma omp simd 18 // CHECK: store i32 0, i32* [[OMP_IV:%[^,]+]] 19 20 // CHECK: [[IV:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID:[0-9]+]] 21 // CHECK-NEXT: [[CMP:%.+]] = icmp slt i32 [[IV]], 6 22 // CHECK-NEXT: br i1 [[CMP]], label %[[SIMPLE_LOOP1_BODY:.+]], label %[[SIMPLE_LOOP1_END:[^,]+]] 23 for (int i = 3; i < 32; i += 5) { 24 // CHECK: [[SIMPLE_LOOP1_BODY]] 25 // Start of body: calculate i from IV: 26 // CHECK: [[IV1_1:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID]] 27 // CHECK: [[CALC_I_1:%.+]] = mul nsw i32 [[IV1_1]], 5 28 // CHECK-NEXT: [[CALC_I_2:%.+]] = add nsw i32 3, [[CALC_I_1]] 29 // CHECK-NEXT: store i32 [[CALC_I_2]], i32* [[LC_I:.+]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID]] 30 // ... loop body ... 31 // End of body: store into a[i]: 32 // CHECK: store float [[RESULT:%.+]], float* {{%.+}}{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID]] 33 a[i] = b[i] * c[i] * d[i]; 34 // CHECK: [[IV1_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID]] 35 // CHECK-NEXT: [[ADD1_2:%.+]] = add nsw i32 [[IV1_2]], 1 36 // CHECK-NEXT: store i32 [[ADD1_2]], i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP1_ID]] 37 // br label %{{.+}}, !llvm.loop !{{.+}} 38 } 39 // CHECK: [[SIMPLE_LOOP1_END]] 40 41 long long k = get_val(); 42 43 #pragma omp simd linear(k : 3) 44 // CHECK: [[K0:%.+]] = call {{.*}}i64 @{{.*}}get_val 45 // CHECK-NEXT: store i64 [[K0]], i64* [[K_VAR:%[^,]+]] 46 // CHECK: store i32 0, i32* [[OMP_IV2:%[^,]+]] 47 // CHECK: [[K0LOAD:%.+]] = load i64, i64* [[K_VAR]] 48 // CHECK-NEXT: store i64 [[K0LOAD]], i64* [[LIN0:%[^,]+]] 49 50 // CHECK: [[IV2:%.+]] = load i32, i32* [[OMP_IV2]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID:[0-9]+]] 51 // CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV2]], 9 52 // CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP2_BODY:.+]], label %[[SIMPLE_LOOP2_END:[^,]+]] 53 for (int i = 10; i > 1; i--) { 54 // CHECK: [[SIMPLE_LOOP2_BODY]] 55 // Start of body: calculate i from IV: 56 // CHECK: [[IV2_0:%.+]] = load i32, i32* [[OMP_IV2]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 57 // FIXME: It is interesting, why the following "mul 1" was not constant folded? 58 // CHECK-NEXT: [[IV2_1:%.+]] = mul nsw i32 [[IV2_0]], 1 59 // CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV2_1]] 60 // CHECK-NEXT: store i32 [[LC_I_1]], i32* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 61 // 62 // CHECK-NEXT: [[LIN0_1:%.+]] = load i64, i64* [[LIN0]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 63 // CHECK-NEXT: [[IV2_2:%.+]] = load i32, i32* [[OMP_IV2]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 64 // CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV2_2]], 3 65 // CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64 66 // CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]] 67 // Update of the privatized version of linear variable! 68 // CHECK-NEXT: store i64 [[LIN_ADD1]], i64* [[K_PRIVATIZED:%[^,]+]] 69 a[k]++; 70 k = k + 3; 71 // CHECK: [[IV2_2:%.+]] = load i32, i32* [[OMP_IV2]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 72 // CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV2_2]], 1 73 // CHECK-NEXT: store i32 [[ADD2_2]], i32* [[OMP_IV2]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP2_ID]] 74 // br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP2_ID]] 75 } 76 // CHECK: [[SIMPLE_LOOP2_END]] 77 // 78 // Update linear vars after loop, as the loop was operating on a private version. 79 // CHECK: [[LIN0_2:%.+]] = load i64, i64* [[LIN0]] 80 // CHECK-NEXT: [[LIN_ADD2:%.+]] = add nsw i64 [[LIN0_2]], 27 81 // CHECK-NEXT: store i64 [[LIN_ADD2]], i64* [[K_VAR]] 82 // 83 84 int lin = 12; 85 #pragma omp simd linear(lin : get_val()), linear(g_ptr) 86 87 // Init linear private var. 88 // CHECK: store i32 12, i32* [[LIN_VAR:%[^,]+]] 89 // CHECK: store i64 0, i64* [[OMP_IV3:%[^,]+]] 90 91 // CHECK: [[LIN_LOAD:%.+]] = load i32, i32* [[LIN_VAR]] 92 // CHECK-NEXT: store i32 [[LIN_LOAD]], i32* [[LIN_START:%[^,]+]] 93 // Remember linear step. 94 // CHECK: [[CALL_VAL:%.+]] = invoke 95 // CHECK: store i64 [[CALL_VAL]], i64* [[LIN_STEP:%[^,]+]] 96 97 // CHECK: [[GLIN_LOAD:%.+]] = load double*, double** [[GLIN_VAR:@[^,]+]] 98 // CHECK-NEXT: store double* [[GLIN_LOAD]], double** [[GLIN_START:%[^,]+]] 99 100 // CHECK: [[IV3:%.+]] = load i64, i64* [[OMP_IV3]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID:[0-9]+]] 101 // CHECK-NEXT: [[CMP3:%.+]] = icmp ult i64 [[IV3]], 4 102 // CHECK-NEXT: br i1 [[CMP3]], label %[[SIMPLE_LOOP3_BODY:.+]], label %[[SIMPLE_LOOP3_END:[^,]+]] 103 for (unsigned long long it = 2000; it >= 600; it-=400) { 104 // CHECK: [[SIMPLE_LOOP3_BODY]] 105 // Start of body: calculate it from IV: 106 // CHECK: [[IV3_0:%.+]] = load i64, i64* [[OMP_IV3]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 107 // CHECK-NEXT: [[LC_IT_1:%.+]] = mul i64 [[IV3_0]], 400 108 // CHECK-NEXT: [[LC_IT_2:%.+]] = sub i64 2000, [[LC_IT_1]] 109 // CHECK-NEXT: store i64 [[LC_IT_2]], i64* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 110 // 111 // Linear start and step are used to calculate current value of the linear variable. 112 // CHECK: [[LINSTART:.+]] = load i32, i32* [[LIN_START]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 113 // CHECK: [[LINSTEP:.+]] = load i64, i64* [[LIN_STEP]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 114 // CHECK-NOT: store i32 {{.+}}, i32* [[LIN_VAR]],{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 115 // CHECK: [[GLINSTART:.+]] = load double*, double** [[GLIN_START]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 116 // CHECK-NEXT: [[IV3_1:%.+]] = load i64, i64* [[OMP_IV3]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 117 // CHECK-NEXT: [[MUL:%.+]] = mul i64 [[IV3_1]], 1 118 // CHECK: [[GEP:%.+]] = getelementptr{{.*}}[[GLINSTART]] 119 // CHECK-NEXT: store double* [[GEP]], double** [[G_PTR_CUR:%[^,]+]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 120 *g_ptr++ = 0.0; 121 // CHECK: [[GEP_VAL:%.+]] = load double{{.*}}[[G_PTR_CUR]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 122 // CHECK: store double{{.*}}[[GEP_VAL]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 123 a[it + lin]++; 124 // CHECK: [[FLT_INC:%.+]] = fadd float 125 // CHECK-NEXT: store float [[FLT_INC]],{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 126 // CHECK: [[IV3_2:%.+]] = load i64, i64* [[OMP_IV3]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 127 // CHECK-NEXT: [[ADD3_2:%.+]] = add i64 [[IV3_2]], 1 128 // CHECK-NEXT: store i64 [[ADD3_2]], i64* [[OMP_IV3]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP3_ID]] 129 } 130 // CHECK: [[SIMPLE_LOOP3_END]] 131 // 132 // Linear start and step are used to calculate final value of the linear variables. 133 // CHECK: [[LINSTART:.+]] = load i32, i32* [[LIN_START]] 134 // CHECK: [[LINSTEP:.+]] = load i64, i64* [[LIN_STEP]] 135 // CHECK: store i32 {{.+}}, i32* [[LIN_VAR]], 136 // CHECK: [[GLINSTART:.+]] = load double*, double** [[GLIN_START]] 137 // CHECK: store double* {{.*}}[[GLIN_VAR]] 138 139 #pragma omp simd 140 // CHECK: store i32 0, i32* [[OMP_IV4:%[^,]+]] 141 142 // CHECK: [[IV4:%.+]] = load i32, i32* [[OMP_IV4]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP4_ID:[0-9]+]] 143 // CHECK-NEXT: [[CMP4:%.+]] = icmp slt i32 [[IV4]], 4 144 // CHECK-NEXT: br i1 [[CMP4]], label %[[SIMPLE_LOOP4_BODY:.+]], label %[[SIMPLE_LOOP4_END:[^,]+]] 145 for (short it = 6; it <= 20; it-=-4) { 146 // CHECK: [[SIMPLE_LOOP4_BODY]] 147 // Start of body: calculate it from IV: 148 // CHECK: [[IV4_0:%.+]] = load i32, i32* [[OMP_IV4]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP4_ID]] 149 // CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i32 [[IV4_0]], 4 150 // CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i32 6, [[LC_IT_1]] 151 // CHECK-NEXT: [[LC_IT_3:%.+]] = trunc i32 [[LC_IT_2]] to i16 152 // CHECK-NEXT: store i16 [[LC_IT_3]], i16* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP4_ID]] 153 154 // CHECK: [[IV4_2:%.+]] = load i32, i32* [[OMP_IV4]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP4_ID]] 155 // CHECK-NEXT: [[ADD4_2:%.+]] = add nsw i32 [[IV4_2]], 1 156 // CHECK-NEXT: store i32 [[ADD4_2]], i32* [[OMP_IV4]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP4_ID]] 157 } 158 // CHECK: [[SIMPLE_LOOP4_END]] 159 160 #pragma omp simd 161 // CHECK: store i32 0, i32* [[OMP_IV5:%[^,]+]] 162 163 // CHECK: [[IV5:%.+]] = load i32, i32* [[OMP_IV5]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP5_ID:[0-9]+]] 164 // CHECK-NEXT: [[CMP5:%.+]] = icmp slt i32 [[IV5]], 26 165 // CHECK-NEXT: br i1 [[CMP5]], label %[[SIMPLE_LOOP5_BODY:.+]], label %[[SIMPLE_LOOP5_END:[^,]+]] 166 for (unsigned char it = 'z'; it >= 'a'; it+=-1) { 167 // CHECK: [[SIMPLE_LOOP5_BODY]] 168 // Start of body: calculate it from IV: 169 // CHECK: [[IV5_0:%.+]] = load i32, i32* [[OMP_IV5]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP5_ID]] 170 // CHECK-NEXT: [[IV5_1:%.+]] = mul nsw i32 [[IV5_0]], 1 171 // CHECK-NEXT: [[LC_IT_1:%.+]] = sub nsw i32 122, [[IV5_1]] 172 // CHECK-NEXT: [[LC_IT_2:%.+]] = trunc i32 [[LC_IT_1]] to i8 173 // CHECK-NEXT: store i8 [[LC_IT_2]], i8* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP5_ID]] 174 175 // CHECK: [[IV5_2:%.+]] = load i32, i32* [[OMP_IV5]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP5_ID]] 176 // CHECK-NEXT: [[ADD5_2:%.+]] = add nsw i32 [[IV5_2]], 1 177 // CHECK-NEXT: store i32 [[ADD5_2]], i32* [[OMP_IV5]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP5_ID]] 178 } 179 // CHECK: [[SIMPLE_LOOP5_END]] 180 181 // CHECK-NOT: mul i32 %{{.+}}, 10 182 #pragma omp simd 183 for (unsigned i=100; i<10; i+=10) { 184 } 185 186 int A; 187 // CHECK: store i32 -1, i32* [[A:%.+]], 188 A = -1; 189 #pragma omp simd lastprivate(A) 190 // CHECK: store i64 0, i64* [[OMP_IV7:%[^,]+]] 191 // CHECK: br label %[[SIMD_LOOP7_COND:[^,]+]] 192 // CHECK: [[SIMD_LOOP7_COND]] 193 // CHECK-NEXT: [[IV7:%.+]] = load i64, i64* [[OMP_IV7]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID:[0-9]+]] 194 // CHECK-NEXT: [[CMP7:%.+]] = icmp slt i64 [[IV7]], 7 195 // CHECK-NEXT: br i1 [[CMP7]], label %[[SIMPLE_LOOP7_BODY:.+]], label %[[SIMPLE_LOOP7_END:[^,]+]] 196 for (long long i = -10; i < 10; i += 3) { 197 // CHECK: [[SIMPLE_LOOP7_BODY]] 198 // Start of body: calculate i from IV: 199 // CHECK: [[IV7_0:%.+]] = load i64, i64* [[OMP_IV7]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 200 // CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i64 [[IV7_0]], 3 201 // CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i64 -10, [[LC_IT_1]] 202 // CHECK-NEXT: store i64 [[LC_IT_2]], i64* [[LC:%[^,]+]],{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 203 // CHECK-NEXT: [[LC_VAL:%.+]] = load i64, i64* [[LC]]{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 204 // CHECK-NEXT: [[CONV:%.+]] = trunc i64 [[LC_VAL]] to i32 205 // CHECK-NEXT: store i32 [[CONV]], i32* [[A_PRIV:%[^,]+]],{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 206 A = i; 207 // CHECK: [[IV7_2:%.+]] = load i64, i64* [[OMP_IV7]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 208 // CHECK-NEXT: [[ADD7_2:%.+]] = add nsw i64 [[IV7_2]], 1 209 // CHECK-NEXT: store i64 [[ADD7_2]], i64* [[OMP_IV7]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP7_ID]] 210 } 211 // CHECK: [[SIMPLE_LOOP7_END]] 212 // CHECK-NEXT: store i64 11, i64* 213 // CHECK-NEXT: [[A_PRIV_VAL:%.+]] = load i32, i32* [[A_PRIV]], 214 // CHECK-NEXT: store i32 [[A_PRIV_VAL]], i32* [[A]], 215 int R; 216 // CHECK: store i32 -1, i32* [[R:%[^,]+]], 217 R = -1; 218 // CHECK: store i64 0, i64* [[OMP_IV8:%[^,]+]], 219 // CHECK: store i32 1, i32* [[R_PRIV:%[^,]+]], 220 #pragma omp simd reduction(*:R) 221 // CHECK: br label %[[SIMD_LOOP8_COND:[^,]+]] 222 // CHECK: [[SIMD_LOOP8_COND]] 223 // CHECK-NEXT: [[IV8:%.+]] = load i64, i64* [[OMP_IV8]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID:[0-9]+]] 224 // CHECK-NEXT: [[CMP8:%.+]] = icmp slt i64 [[IV8]], 7 225 // CHECK-NEXT: br i1 [[CMP8]], label %[[SIMPLE_LOOP8_BODY:.+]], label %[[SIMPLE_LOOP8_END:[^,]+]] 226 for (long long i = -10; i < 10; i += 3) { 227 // CHECK: [[SIMPLE_LOOP8_BODY]] 228 // Start of body: calculate i from IV: 229 // CHECK: [[IV8_0:%.+]] = load i64, i64* [[OMP_IV8]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 230 // CHECK-NEXT: [[LC_IT_1:%.+]] = mul nsw i64 [[IV8_0]], 3 231 // CHECK-NEXT: [[LC_IT_2:%.+]] = add nsw i64 -10, [[LC_IT_1]] 232 // CHECK-NEXT: store i64 [[LC_IT_2]], i64* [[LC:%[^,]+]],{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 233 // CHECK-NEXT: [[LC_VAL:%.+]] = load i64, i64* [[LC]]{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 234 // CHECK: store i32 %{{.+}}, i32* [[R_PRIV]],{{.+}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 235 R *= i; 236 // CHECK: [[IV8_2:%.+]] = load i64, i64* [[OMP_IV8]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 237 // CHECK-NEXT: [[ADD8_2:%.+]] = add nsw i64 [[IV8_2]], 1 238 // CHECK-NEXT: store i64 [[ADD8_2]], i64* [[OMP_IV8]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP8_ID]] 239 } 240 // CHECK: [[SIMPLE_LOOP8_END]] 241 // CHECK-DAG: [[R_VAL:%.+]] = load i32, i32* [[R]], 242 // CHECK-DAG: [[R_PRIV_VAL:%.+]] = load i32, i32* [[R_PRIV]], 243 // CHECK: [[RED:%.+]] = mul nsw i32 [[R_VAL]], [[R_PRIV_VAL]] 244 // CHECK-NEXT: store i32 [[RED]], i32* [[R]], 245 // CHECK-NEXT: ret void 246 } 247 248 template <class T, unsigned K> T tfoo(T a) { return a + K; } 249 250 template <typename T, unsigned N> 251 int templ1(T a, T *z) { 252 #pragma omp simd collapse(N) 253 for (int i = 0; i < N * 2; i++) { 254 for (long long j = 0; j < (N + N + N + N); j += 2) { 255 z[i + j] = a + tfoo<T, N>(i + j); 256 } 257 } 258 return 0; 259 } 260 261 // Instatiation templ1<float,2> 262 // CHECK-LABEL: define {{.*i32}} @{{.*}}templ1{{.*}}(float {{.+}}, float* {{.+}}) 263 // CHECK: store i64 0, i64* [[T1_OMP_IV:[^,]+]] 264 // ... 265 // CHECK: [[IV:%.+]] = load i64, i64* [[T1_OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID:[0-9]+]] 266 // CHECK-NEXT: [[CMP1:%.+]] = icmp slt i64 [[IV]], 16 267 // CHECK-NEXT: br i1 [[CMP1]], label %[[T1_BODY:.+]], label %[[T1_END:[^,]+]] 268 // CHECK: [[T1_BODY]] 269 // Loop counters i and j updates: 270 // CHECK: [[IV1:%.+]] = load i64, i64* [[T1_OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 271 // CHECK-NEXT: [[I_1:%.+]] = sdiv i64 [[IV1]], 4 272 // CHECK-NEXT: [[I_1_MUL1:%.+]] = mul nsw i64 [[I_1]], 1 273 // CHECK-NEXT: [[I_1_ADD0:%.+]] = add nsw i64 0, [[I_1_MUL1]] 274 // CHECK-NEXT: [[I_2:%.+]] = trunc i64 [[I_1_ADD0]] to i32 275 // CHECK-NEXT: store i32 [[I_2]], i32* {{%.+}}{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 276 // CHECK: [[IV2:%.+]] = load i64, i64* [[T1_OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 277 // CHECK-NEXT: [[J_1:%.+]] = srem i64 [[IV2]], 4 278 // CHECK-NEXT: [[J_2:%.+]] = mul nsw i64 [[J_1]], 2 279 // CHECK-NEXT: [[J_2_ADD0:%.+]] = add nsw i64 0, [[J_2]] 280 // CHECK-NEXT: store i64 [[J_2_ADD0]], i64* {{%.+}}{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 281 // simd.for.inc: 282 // CHECK: [[IV3:%.+]] = load i64, i64* [[T1_OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 283 // CHECK-NEXT: [[INC:%.+]] = add nsw i64 [[IV3]], 1 284 // CHECK-NEXT: store i64 [[INC]], i64* [[T1_OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[T1_ID]] 285 // CHECK-NEXT: br label {{%.+}} 286 // CHECK: [[T1_END]] 287 // CHECK: ret i32 0 288 // 289 void inst_templ1() { 290 float a; 291 float z[100]; 292 templ1<float,2> (a, z); 293 } 294 295 296 typedef int MyIdx; 297 298 class IterDouble { 299 double *Ptr; 300 public: 301 IterDouble operator++ () const { 302 IterDouble n; 303 n.Ptr = Ptr + 1; 304 return n; 305 } 306 bool operator < (const IterDouble &that) const { 307 return Ptr < that.Ptr; 308 } 309 double & operator *() const { 310 return *Ptr; 311 } 312 MyIdx operator - (const IterDouble &that) const { 313 return (MyIdx) (Ptr - that.Ptr); 314 } 315 IterDouble operator + (int Delta) { 316 IterDouble re; 317 re.Ptr = Ptr + Delta; 318 return re; 319 } 320 321 ///~IterDouble() {} 322 }; 323 324 // CHECK-LABEL: define {{.*void}} @{{.*}}iter_simple{{.*}} 325 void iter_simple(IterDouble ia, IterDouble ib, IterDouble ic) { 326 // 327 // Calculate number of iterations before the loop body. 328 // CHECK: [[DIFF1:%.+]] = invoke {{.*}}i32 @{{.*}}IterDouble{{.*}} 329 // CHECK: [[DIFF2:%.+]] = sub nsw i32 [[DIFF1]], 1 330 // CHECK-NEXT: [[DIFF3:%.+]] = add nsw i32 [[DIFF2]], 1 331 // CHECK-NEXT: [[DIFF4:%.+]] = sdiv i32 [[DIFF3]], 1 332 // CHECK-NEXT: [[DIFF5:%.+]] = sub nsw i32 [[DIFF4]], 1 333 // CHECK-NEXT: store i32 [[DIFF5]], i32* [[OMP_LAST_IT:%[^,]+]]{{.+}} 334 // CHECK: store i32 0, i32* [[IT_OMP_IV:%[^,]+]] 335 #pragma omp simd 336 337 // CHECK: [[IV:%.+]] = load i32, i32* [[IT_OMP_IV]]{{.+}} !llvm.mem.parallel_loop_access ![[ITER_LOOP_ID:[0-9]+]] 338 // CHECK-NEXT: [[LAST_IT:%.+]] = load i32, i32* [[OMP_LAST_IT]]{{.+}}!llvm.mem.parallel_loop_access ![[ITER_LOOP_ID]] 339 // CHECK-NEXT: [[NUM_IT:%.+]] = add nsw i32 [[LAST_IT]], 1 340 // CHECK-NEXT: [[CMP:%.+]] = icmp slt i32 [[IV]], [[NUM_IT]] 341 // CHECK-NEXT: br i1 [[CMP]], label %[[IT_BODY:[^,]+]], label %[[IT_END:[^,]+]] 342 for (IterDouble i = ia; i < ib; ++i) { 343 // CHECK: [[IT_BODY]] 344 // Start of body: calculate i from index: 345 // CHECK: [[IV1:%.+]] = load i32, i32* [[IT_OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[ITER_LOOP_ID]] 346 // Call of operator+ (i, IV). 347 // CHECK: {{%.+}} = invoke {{.+}} @{{.*}}IterDouble{{.*}} 348 // ... loop body ... 349 *i = *ic * 0.5; 350 // Float multiply and save result. 351 // CHECK: [[MULR:%.+]] = fmul double {{%.+}}, 5.000000e-01 352 // CHECK-NEXT: invoke {{.+}} @{{.*}}IterDouble{{.*}} 353 // CHECK: store double [[MULR:%.+]], double* [[RESULT_ADDR:%.+]], !llvm.mem.parallel_loop_access ![[ITER_LOOP_ID]] 354 ++ic; 355 // 356 // CHECK: [[IV2:%.+]] = load i32, i32* [[IT_OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[ITER_LOOP_ID]] 357 // CHECK-NEXT: [[ADD2:%.+]] = add nsw i32 [[IV2]], 1 358 // CHECK-NEXT: store i32 [[ADD2]], i32* [[IT_OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[ITER_LOOP_ID]] 359 // br label %{{.*}}, !llvm.loop ![[ITER_LOOP_ID]] 360 } 361 // CHECK: [[IT_END]] 362 // CHECK: ret void 363 } 364 365 366 // CHECK-LABEL: define {{.*void}} @{{.*}}collapsed{{.*}} 367 void collapsed(float *a, float *b, float *c, float *d) { 368 int i; // outer loop counter 369 unsigned j; // middle loop couter, leads to unsigned icmp in loop header. 370 // k declared in the loop init below 371 short l; // inner loop counter 372 // CHECK: store i32 0, i32* [[OMP_IV:[^,]+]] 373 // 374 #pragma omp simd collapse(4) 375 376 // CHECK: [[IV:%.+]] = load i32, i32* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID:[0-9]+]] 377 // CHECK-NEXT: [[CMP:%.+]] = icmp ult i32 [[IV]], 120 378 // CHECK-NEXT: br i1 [[CMP]], label %[[COLL1_BODY:[^,]+]], label %[[COLL1_END:[^,]+]] 379 for (i = 1; i < 3; i++) // 2 iterations 380 for (j = 2u; j < 5u; j++) //3 iterations 381 for (int k = 3; k <= 6; k++) // 4 iterations 382 for (l = 4; l < 9; ++l) // 5 iterations 383 { 384 // CHECK: [[COLL1_BODY]] 385 // Start of body: calculate i from index: 386 // CHECK: [[IV1:%.+]] = load i32, i32* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 387 // Calculation of the loop counters values. 388 // CHECK: [[CALC_I_1:%.+]] = udiv i32 [[IV1]], 60 389 // CHECK-NEXT: [[CALC_I_1_MUL1:%.+]] = mul i32 [[CALC_I_1]], 1 390 // CHECK-NEXT: [[CALC_I_2:%.+]] = add i32 1, [[CALC_I_1_MUL1]] 391 // CHECK-NEXT: store i32 [[CALC_I_2]], i32* [[LC_I:.+]] 392 // CHECK: [[IV1_2:%.+]] = load i32, i32* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 393 // CHECK-NEXT: [[CALC_J_1:%.+]] = udiv i32 [[IV1_2]], 20 394 // CHECK-NEXT: [[CALC_J_2:%.+]] = urem i32 [[CALC_J_1]], 3 395 // CHECK-NEXT: [[CALC_J_2_MUL1:%.+]] = mul i32 [[CALC_J_2]], 1 396 // CHECK-NEXT: [[CALC_J_3:%.+]] = add i32 2, [[CALC_J_2_MUL1]] 397 // CHECK-NEXT: store i32 [[CALC_J_3]], i32* [[LC_J:.+]] 398 // CHECK: [[IV1_3:%.+]] = load i32, i32* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 399 // CHECK-NEXT: [[CALC_K_1:%.+]] = udiv i32 [[IV1_3]], 5 400 // CHECK-NEXT: [[CALC_K_2:%.+]] = urem i32 [[CALC_K_1]], 4 401 // CHECK-NEXT: [[CALC_K_2_MUL1:%.+]] = mul i32 [[CALC_K_2]], 1 402 // CHECK-NEXT: [[CALC_K_3:%.+]] = add i32 3, [[CALC_K_2_MUL1]] 403 // CHECK-NEXT: store i32 [[CALC_K_3]], i32* [[LC_K:.+]] 404 // CHECK: [[IV1_4:%.+]] = load i32, i32* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 405 // CHECK-NEXT: [[CALC_L_1:%.+]] = urem i32 [[IV1_4]], 5 406 // CHECK-NEXT: [[CALC_L_1_MUL1:%.+]] = mul i32 [[CALC_L_1]], 1 407 // CHECK-NEXT: [[CALC_L_2:%.+]] = add i32 4, [[CALC_L_1_MUL1]] 408 // CHECK-NEXT: [[CALC_L_3:%.+]] = trunc i32 [[CALC_L_2]] to i16 409 // CHECK-NEXT: store i16 [[CALC_L_3]], i16* [[LC_L:.+]] 410 // ... loop body ... 411 // End of body: store into a[i]: 412 // CHECK: store float [[RESULT:%.+]], float* [[RESULT_ADDR:%.+]]{{.+}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 413 float res = b[j] * c[k]; 414 a[i] = res * d[l]; 415 // CHECK: [[IV2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 416 // CHECK-NEXT: [[ADD2:%.+]] = add i32 [[IV2]], 1 417 // CHECK-NEXT: store i32 [[ADD2]], i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[COLL1_LOOP_ID]] 418 // br label %{{[^,]+}}, !llvm.loop ![[COLL1_LOOP_ID]] 419 // CHECK: [[COLL1_END]] 420 } 421 // i,j,l are updated; k is not updated. 422 // CHECK: store i32 3, i32* 423 // CHECK-NEXT: store i32 5, i32* 424 // CHECK-NEXT: store i32 7, i32* 425 // CHECK-NEXT: store i16 9, i16* 426 // CHECK: ret void 427 } 428 429 extern char foo(); 430 extern double globalfloat; 431 432 // CHECK-LABEL: define {{.*void}} @{{.*}}widened{{.*}} 433 void widened(float *a, float *b, float *c, float *d) { 434 int i; // outer loop counter 435 short j; // inner loop counter 436 globalfloat = 1.0; 437 int localint = 1; 438 // CHECK: store double {{.+}}, double* [[GLOBALFLOAT:@.+]] 439 // Counter is widened to 64 bits. 440 // CHECK: store i64 0, i64* [[OMP_IV:[^,]+]] 441 // 442 #pragma omp simd collapse(2) private(globalfloat, localint) 443 444 // CHECK: [[IV:%.+]] = load i64, i64* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID:[0-9]+]] 445 // CHECK-NEXT: [[LI:%.+]] = load i64, i64* [[OMP_LI:%[^,]+]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 446 // CHECK-NEXT: [[NUMIT:%.+]] = add nsw i64 [[LI]], 1 447 // CHECK-NEXT: [[CMP:%.+]] = icmp slt i64 [[IV]], [[NUMIT]] 448 // CHECK-NEXT: br i1 [[CMP]], label %[[WIDE1_BODY:[^,]+]], label %[[WIDE1_END:[^,]+]] 449 for (i = 1; i < 3; i++) // 2 iterations 450 for (j = 0; j < foo(); j++) // foo() iterations 451 { 452 // CHECK: [[WIDE1_BODY]] 453 // Start of body: calculate i from index: 454 // CHECK: [[IV1:%.+]] = load i64, i64* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 455 // Calculation of the loop counters values... 456 // CHECK: store i32 {{[^,]+}}, i32* [[LC_I:.+]] 457 // CHECK: [[IV1_2:%.+]] = load i64, i64* [[OMP_IV]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 458 // CHECK: store i16 {{[^,]+}}, i16* [[LC_J:.+]] 459 // ... loop body ... 460 // 461 // Here we expect store into private double var, not global 462 // CHECK-NOT: store double {{.+}}, double* [[GLOBALFLOAT]] 463 globalfloat = (float)j/i; 464 float res = b[j] * c[j]; 465 // Store into a[i]: 466 // CHECK: store float [[RESULT:%.+]], float* [[RESULT_ADDR:%.+]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 467 a[i] = res * d[i]; 468 // Then there's a store into private var localint: 469 // CHECK: store i32 {{.+}}, i32* [[LOCALINT:%[^,]+]]{{.+}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 470 localint = (int)j; 471 // CHECK: [[IV2:%.+]] = load i64, i64* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 472 // CHECK-NEXT: [[ADD2:%.+]] = add nsw i64 [[IV2]], 1 473 // CHECK-NEXT: store i64 [[ADD2]], i64* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[WIDE1_LOOP_ID]] 474 // 475 // br label %{{[^,]+}}, !llvm.loop ![[WIDE1_LOOP_ID]] 476 // CHECK: [[WIDE1_END]] 477 } 478 // i,j are updated. 479 // CHECK: store i32 3, i32* [[I:%[^,]+]] 480 // CHECK: store i16 481 // 482 // Here we expect store into original localint, not its privatized version. 483 // CHECK-NOT: store i32 {{.+}}, i32* [[LOCALINT]] 484 localint = (int)j; 485 // CHECK: ret void 486 } 487 488 // CHECK-LABEL: define {{.*void}} @{{.*}}linear{{.*}}(float* {{.+}}) 489 void linear(float *a) { 490 // CHECK: [[VAL_ADDR:%.+]] = alloca i64, 491 // CHECK: [[K_ADDR:%.+]] = alloca i64*, 492 long long val = 0; 493 long long &k = val; 494 495 #pragma omp simd linear(k : 3) 496 // CHECK: store i64* [[VAL_ADDR]], i64** [[K_ADDR]], 497 // CHECK: [[VAL_REF:%.+]] = load i64*, i64** [[K_ADDR]], 498 // CHECK: store i64* [[VAL_REF]], i64** [[K_ADDR_REF:%.+]], 499 // CHECK: store i32 0, i32* [[OMP_IV:%[^,]+]] 500 // CHECK: [[K_REF:%.+]] = load i64*, i64** [[K_ADDR_REF]], 501 // CHECK: [[K0LOAD:%.+]] = load i64, i64* [[K_REF]] 502 // CHECK-NEXT: store i64 [[K0LOAD]], i64* [[LIN0:%[^,]+]] 503 504 // CHECK: [[IV:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID:[0-9]+]] 505 // CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9 506 // CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]] 507 for (int i = 10; i > 1; i--) { 508 // CHECK: [[SIMPLE_LOOP_BODY]] 509 // Start of body: calculate i from IV: 510 // CHECK: [[IV_0:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 511 // FIXME: It is interesting, why the following "mul 1" was not constant folded? 512 // CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1 513 // CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]] 514 // CHECK-NEXT: store i32 [[LC_I_1]], i32* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 515 // 516 // CHECK-NEXT: [[LIN0_1:%.+]] = load i64, i64* [[LIN0]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 517 // CHECK-NEXT: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 518 // CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3 519 // CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64 520 // CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]] 521 // Update of the privatized version of linear variable! 522 // CHECK-NEXT: store i64 [[LIN_ADD1]], i64* [[K_PRIVATIZED:%[^,]+]] 523 a[k]++; 524 k = k + 3; 525 // CHECK: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 526 // CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1 527 // CHECK-NEXT: store i32 [[ADD2_2]], i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 528 // br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]] 529 } 530 // CHECK: [[SIMPLE_LOOP_END]] 531 // 532 // Update linear vars after loop, as the loop was operating on a private version. 533 // CHECK: [[K_REF:%.+]] = load i64*, i64** [[K_ADDR_REF]], 534 // CHECK: store i64* [[K_REF]], i64** [[K_PRIV_REF:%.+]], 535 // CHECK: [[LIN0_2:%.+]] = load i64, i64* [[LIN0]] 536 // CHECK-NEXT: [[LIN_ADD2:%.+]] = add nsw i64 [[LIN0_2]], 27 537 // CHECK-NEXT: [[K_REF:%.+]] = load i64*, i64** [[K_PRIV_REF]], 538 // CHECK-NEXT: store i64 [[LIN_ADD2]], i64* [[K_REF]] 539 // 540 541 #pragma omp simd linear(val(k) : 3) 542 // CHECK: [[VAL_REF:%.+]] = load i64*, i64** [[K_ADDR]], 543 // CHECK: store i64* [[VAL_REF]], i64** [[K_ADDR_REF:%.+]], 544 // CHECK: store i32 0, i32* [[OMP_IV:%[^,]+]] 545 // CHECK: [[K_REF:%.+]] = load i64*, i64** [[K_ADDR_REF]], 546 // CHECK: [[K0LOAD:%.+]] = load i64, i64* [[K_REF]] 547 // CHECK-NEXT: store i64 [[K0LOAD]], i64* [[LIN0:%[^,]+]] 548 549 // CHECK: [[IV:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID:[0-9]+]] 550 // CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9 551 // CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]] 552 for (int i = 10; i > 1; i--) { 553 // CHECK: [[SIMPLE_LOOP_BODY]] 554 // Start of body: calculate i from IV: 555 // CHECK: [[IV_0:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 556 // FIXME: It is interesting, why the following "mul 1" was not constant folded? 557 // CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1 558 // CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]] 559 // CHECK-NEXT: store i32 [[LC_I_1]], i32* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 560 // 561 // CHECK-NEXT: [[LIN0_1:%.+]] = load i64, i64* [[LIN0]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 562 // CHECK-NEXT: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 563 // CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3 564 // CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64 565 // CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]] 566 // Update of the privatized version of linear variable! 567 // CHECK-NEXT: store i64 [[LIN_ADD1]], i64* [[K_PRIVATIZED:%[^,]+]] 568 a[k]++; 569 k = k + 3; 570 // CHECK: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 571 // CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1 572 // CHECK-NEXT: store i32 [[ADD2_2]], i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 573 // br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]] 574 } 575 // CHECK: [[SIMPLE_LOOP_END]] 576 // 577 // Update linear vars after loop, as the loop was operating on a private version. 578 // CHECK: [[K_REF:%.+]] = load i64*, i64** [[K_ADDR_REF]], 579 // CHECK: store i64* [[K_REF]], i64** [[K_PRIV_REF:%.+]], 580 // CHECK: [[LIN0_2:%.+]] = load i64, i64* [[LIN0]] 581 // CHECK-NEXT: [[LIN_ADD2:%.+]] = add nsw i64 [[LIN0_2]], 27 582 // CHECK-NEXT: [[K_REF:%.+]] = load i64*, i64** [[K_PRIV_REF]], 583 // CHECK-NEXT: store i64 [[LIN_ADD2]], i64* [[K_REF]] 584 // 585 #pragma omp simd linear(uval(k) : 3) 586 // CHECK: store i32 0, i32* [[OMP_IV:%[^,]+]] 587 // CHECK: [[K0LOAD:%.+]] = load i64, i64* [[VAL_ADDR]] 588 // CHECK-NEXT: store i64 [[K0LOAD]], i64* [[LIN0:%[^,]+]] 589 590 // CHECK: [[IV:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID:[0-9]+]] 591 // CHECK-NEXT: [[CMP2:%.+]] = icmp slt i32 [[IV]], 9 592 // CHECK-NEXT: br i1 [[CMP2]], label %[[SIMPLE_LOOP_BODY:.+]], label %[[SIMPLE_LOOP_END:[^,]+]] 593 for (int i = 10; i > 1; i--) { 594 // CHECK: [[SIMPLE_LOOP_BODY]] 595 // Start of body: calculate i from IV: 596 // CHECK: [[IV_0:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 597 // FIXME: It is interesting, why the following "mul 1" was not constant folded? 598 // CHECK-NEXT: [[IV_1:%.+]] = mul nsw i32 [[IV_0]], 1 599 // CHECK-NEXT: [[LC_I_1:%.+]] = sub nsw i32 10, [[IV_1]] 600 // CHECK-NEXT: store i32 [[LC_I_1]], i32* {{.+}}, !llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 601 // 602 // CHECK-NEXT: [[LIN0_1:%.+]] = load i64, i64* [[LIN0]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 603 // CHECK-NEXT: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 604 // CHECK-NEXT: [[LIN_MUL1:%.+]] = mul nsw i32 [[IV_2]], 3 605 // CHECK-NEXT: [[LIN_EXT1:%.+]] = sext i32 [[LIN_MUL1]] to i64 606 // CHECK-NEXT: [[LIN_ADD1:%.+]] = add nsw i64 [[LIN0_1]], [[LIN_EXT1]] 607 // Update of the privatized version of linear variable! 608 // CHECK-NEXT: store i64 [[LIN_ADD1]], i64* [[K_PRIVATIZED:%[^,]+]] 609 a[k]++; 610 k = k + 3; 611 // CHECK: [[IV_2:%.+]] = load i32, i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 612 // CHECK-NEXT: [[ADD2_2:%.+]] = add nsw i32 [[IV_2]], 1 613 // CHECK-NEXT: store i32 [[ADD2_2]], i32* [[OMP_IV]]{{.*}}!llvm.mem.parallel_loop_access ![[SIMPLE_LOOP_ID]] 614 // br label {{.+}}, !llvm.loop ![[SIMPLE_LOOP_ID]] 615 } 616 // CHECK: [[SIMPLE_LOOP_END]] 617 // 618 // Update linear vars after loop, as the loop was operating on a private version. 619 // CHECK: [[LIN0_2:%.+]] = load i64, i64* [[LIN0]] 620 // CHECK-NEXT: [[LIN_ADD2:%.+]] = add nsw i64 [[LIN0_2]], 27 621 // CHECK-NEXT: store i64 [[LIN_ADD2]], i64* [[VAL_ADDR]] 622 // 623 } 624 625 // TERM_DEBUG-LABEL: bar 626 int bar() {return 0;}; 627 628 // TERM_DEBUG-LABEL: parallel_simd 629 void parallel_simd(float *a) { 630 #pragma omp parallel 631 #pragma omp simd 632 // TERM_DEBUG-NOT: __kmpc_global_thread_num 633 // TERM_DEBUG: invoke i32 {{.*}}bar{{.*}}() 634 // TERM_DEBUG: unwind label %[[TERM_LPAD:.+]], 635 // TERM_DEBUG-NOT: __kmpc_global_thread_num 636 // TERM_DEBUG: [[TERM_LPAD]] 637 // TERM_DEBUG: call void @__clang_call_terminate 638 // TERM_DEBUG: unreachable 639 for (unsigned i = 131071; i <= 2147483647; i += 127) 640 a[i] += bar(); 641 } 642 // TERM_DEBUG: !{{[0-9]+}} = !DILocation(line: [[@LINE-11]], 643 644 // CHECK-LABEL: S8 645 // CHECK: ptrtoint [[SS_TY]]* %{{.+}} to i64 646 // CHECK-NEXT: and i64 %{{.+}}, 15 647 // CHECK-NEXT: icmp eq i64 %{{.+}}, 0 648 // CHECK-NEXT: call void @llvm.assume(i1 649 650 // CHECK: ptrtoint [[SS_TY]]* %{{.+}} to i64 651 // CHECK-NEXT: and i64 %{{.+}}, 7 652 // CHECK-NEXT: icmp eq i64 %{{.+}}, 0 653 // CHECK-NEXT: call void @llvm.assume(i1 654 655 // CHECK: ptrtoint [[SS_TY]]* %{{.+}} to i64 656 // CHECK-NEXT: and i64 %{{.+}}, 15 657 // CHECK-NEXT: icmp eq i64 %{{.+}}, 0 658 // CHECK-NEXT: call void @llvm.assume(i1 659 660 // CHECK: ptrtoint [[SS_TY]]* %{{.+}} to i64 661 // CHECK-NEXT: and i64 %{{.+}}, 3 662 // CHECK-NEXT: icmp eq i64 %{{.+}}, 0 663 // CHECK-NEXT: call void @llvm.assume(i1 664 struct SS { 665 SS(): a(0) {} 666 SS(int v) : a(v) {} 667 int a; 668 typedef int type; 669 }; 670 671 template <typename T> 672 class S7 : public T { 673 protected: 674 T *a; 675 T b[2]; 676 S7() : a(0) {} 677 678 public: 679 S7(typename T::type &v) : a((T*)&v) { 680 #pragma omp simd aligned(a) 681 for (int k = 0; k < a->a; ++k) 682 ++this->a->a; 683 #pragma omp simd aligned(this->b : 8) 684 for (int k = 0; k < a->a; ++k) 685 ++a->a; 686 } 687 }; 688 689 class S8 : private IterDouble, public S7<SS> { 690 S8() {} 691 692 public: 693 S8(int v) : S7<SS>(v){ 694 #pragma omp parallel private(a) 695 #pragma omp simd aligned(S7<SS>::a) 696 for (int k = 0; k < a->a; ++k) 697 ++this->a->a; 698 #pragma omp parallel shared(b) 699 #pragma omp simd aligned(this->b: 4) 700 for (int k = 0; k < a->a; ++k) 701 ++a->a; 702 } 703 }; 704 S8 s8(0); 705 706 #endif // HEADER 707 708