#include "pthread_impl.h"
#include <threads.h>

/* future version of C might prescribe a minimum number of tss */
#define __TSS_MAX_MIN 64

#ifndef TSS_MAX
# define TSS_MAX __TSS_MAX_MIN
#endif

#define __LOW_BIT(X) ((X+0U) & -(X+0U))

#define STRINGIFY_(X) #X
#define STRINGIFY(X) STRINGIFY_(X)
#define WARN__(X) _Pragma(#X)
#define WARN_(X) WARN__(message (X))
#define WARN(X) WARN_(X)

#if (TSS_MAX) < __TSS_MAX_MIN
WARN("TSS_MAX " STRINGIFY(TSS_MAX) " replaced by minimal value " STRINGIFY(__TSS_MAX_MIN))
# undef TSS_MAX
# define TSS_MAX __TSS_MAX_MIN
#endif

#if (TSS_MAX) != __LOW_BIT(TSS_MAX)
# error TSS_MAX must be a power of 2
#endif

enum { tss_max = (TSS_MAX), };

static tss_dtor_t dtors[tss_max];

static void nodtor(void *dummy)
{
}

/* This should use _Thread_local which must exist for C11 threads.
   Since musl might get compiled with pre-C11 options we use gcc's
   __thread extension, instead. */
static __thread void * data[tss_max];
/* Thread local data is accounted for with a usage counter */
static __thread tss_t used;

int tss_create(tss_t *tss, tss_dtor_t dtor)
{
	unsigned start = (uintptr_t)&tss / 16 % tss_max;
	if (!dtor) dtor = nodtor;

	for (unsigned i = 0; i < tss_max; ++i) {
		unsigned j = (start+i)%tss_max;
		if (!a_cas_p(dtors+j, 0, (void *)dtor)) {
			*tss = j;
			return thrd_success;
		}
	}
	*tss = -1;
	return thrd_error;
}

void tss_delete(tss_t k)
{
	/* tss_delete has no failure path, so check the argument. */
	if (k < tss_max) dtors[k] = 0;
}

int tss_set(tss_t k, void *x)
{
	if (k >= tss_max || !dtors[k]) return thrd_error;
	if (data[k] != x) {
		/* tss_set can be used to add or remove a tss. compute
		   the difference in usage. */
		used += (!!x - !!data[k]);
		data[k] = x;
	}
	return thrd_success;
}

void * tss_get(tss_t k) {
	return dtors[k] ? data[k] : 0;
}


void __tss_run_dtors(void)
{
	for (unsigned j=0; used && j<TSS_DTOR_ITERATIONS; j++) {
		for (unsigned i=0; i<tss_max; i++) {
			if (data[i]) {
				void *tmp = data[i];
				data[i] = 0;
				--used;
                                // slot i might have been deleted, check
				if (dtors[i]) dtors[i](tmp);
                               	if (!used) return;
			}
		}
	}
}