blob: 403b0394298e12180505ab34f0c39cd323d49b9d [file] [log] [blame]
/* { dg-do compile } */
/* { dg-options "-Wabsolute-value" } */
#include <stdlib.h>
#include <complex.h>
#include <math.h>
void tst_decimal (_Decimal32 *p32, _Decimal64 *p64, _Decimal128 *p128)
{
*p32 = abs(*p32); /* { dg-warning "using integer absolute value function" } */
*p64 = fabs(*p64); /* { dg-warning "using floating-point absolute value function" } */
*p128 = cabsl(*p128); /* { dg-warning "using complex absolute value function" } */
}
void tst_notdecimal (int *pi, double *pd, long double *pld, complex double *pc)
{
*pi = __builtin_fabsd32 (*pi); /* { dg-warning "using decimal floating-point absolute value function" } */
*pd = __builtin_fabsd64 (*pd); /* { dg-warning "using decimal floating-point absolute value function" } */
*pld = __builtin_fabsd64 (*pld); /* { dg-warning "using decimal floating-point absolute value function" } */
*pc = __builtin_fabsd128 (*pc); /* { dg-warning "using decimal floating-point absolute value function" } */
}
void
test_size (_Decimal64 *p64, _Decimal128 *p128)
{
*p64 = __builtin_fabsd32 (*p64); /* { dg-warning "may cause truncation of value" } */
*p128 = __builtin_fabsd64 (*p128); /* { dg-warning "may cause truncation of value" } */
}