1 // RUN: %libomp-cxx-compile-and-run
2 
3 #include <stdio.h>
4 #include <omp.h>
5 
6 #define NT 4
7 #define INIT 10
8 
9 /*
10 The test emulates code generation needed for reduction with task modifier on
11 parallel construct.
12 
13 Note: tasks could just use in_reduction clause, but compiler does not accept
14 this because of bug: it mistakenly requires reduction item to be shared, which
15 is only true for reduction on worksharing and wrong for task reductions.
16 */
17 
18 //------------------------------------------------
19 // OpenMP runtime library routines
20 #ifdef __cplusplus
21 extern "C" {
22 #endif
23 extern void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *item);
24 extern void *__kmpc_task_reduction_modifier_init(void *loc, int gtid, int is_ws,
25                                                  int num, void *data);
26 extern void __kmpc_task_reduction_modifier_fini(void *loc, int gtid, int is_ws);
27 extern int __kmpc_global_thread_num(void *);
28 #ifdef __cplusplus
29 }
30 #endif
31 
32 //------------------------------------------------
33 // Compiler-generated code
34 
35 typedef struct red_input {
36   void *reduce_shar; /**< shared between tasks item to reduce into */
37   size_t reduce_size; /**< size of data item in bytes */
38   // three compiler-generated routines (init, fini are optional):
39   void *reduce_init; /**< data initialization routine (single parameter) */
40   void *reduce_fini; /**< data finalization routine */
41   void *reduce_comb; /**< data combiner routine */
42   unsigned flags; /**< flags for additional info from compiler */
43 } red_input_t;
44 
45 void i_comb(void *lhs, void *rhs) { *(int *)lhs += *(int *)rhs; }
46 
47 int main() {
48   int var = INIT;
49   int i;
50   omp_set_dynamic(0);
51   omp_set_num_threads(NT);
52 #pragma omp parallel private(i)
53 //  #pragma omp for reduction(task,+:var)
54 #pragma omp for reduction(+ : var)
55   for (i = 0; i < NT; ++i) // single iteration per thread
56   {
57     // generated code, which actually should be placed before
58     // loop iterations distribution, but placed here just to show the idea,
59     // and to keep correctness the loop count is equal to number of threads
60     int gtid = __kmpc_global_thread_num(NULL);
61     void *tg; // pointer to taskgroup (optional)
62     red_input_t r_var;
63     r_var.reduce_shar = &var;
64     r_var.reduce_size = sizeof(var);
65     r_var.reduce_init = NULL;
66     r_var.reduce_fini = NULL;
67     r_var.reduce_comb = (void *)&i_comb;
68     tg = __kmpc_task_reduction_modifier_init(
69         NULL, // ident_t loc;
70         gtid,
71         1, // 1 - worksharing construct, 0 - parallel
72         1, // number of reduction objects
73         &r_var // related data
74         );
75     // end of generated code
76     var++;
77 #pragma omp task /*in_reduction(+:var)*/ shared(var)
78     {
79       // emulate task reduction here because of compiler bug:
80       // it mistakenly declines to accept in_reduction because var is private
81       // outside.
82       int gtid = __kmpc_global_thread_num(NULL);
83       int *p_var = (int *)__kmpc_task_reduction_get_th_data(gtid, tg, &var);
84       *p_var += 1;
85     }
86     if (omp_get_thread_num() > 0) {
87 #pragma omp task /*in_reduction(+:var)*/ shared(var)
88       {
89         int gtid = __kmpc_global_thread_num(NULL);
90         int *p_var = (int *)__kmpc_task_reduction_get_th_data(gtid, tg, &var);
91         *p_var += 1;
92       }
93     }
94     // generated code, which actually should be placed after loop completion
95     // but before barrier and before loop reduction. It placed here just to show
96     // the idea,
97     // and to keep correctness the loop count is equal to number of threads
98     __kmpc_task_reduction_modifier_fini(NULL, gtid, 1);
99     // end of generated code
100   }
101   if (var == INIT + NT * 3 - 1) {
102     printf("passed\n");
103     return 0;
104   } else {
105     printf("failed: var = %d (!= %d)\n", var, INIT + NT * 3 - 1);
106     return 1;
107   }
108 }
109