The main idea is to avoid a longjmp and return to the destructor
for the cancelled thread. So, adjust GC and gc_minor to allow
for a NULL continuation.
Handle this by checking if the argument is the primordial thread,
current thread or another thread.
The first two cases remain almost identical to the previous implementation.
To terminate a thread (which is not the caller) we use a pthread key
which contains the thread data. The destructor is set to Cyc_end_thread
and will terminate the thread when pthread_cancel is called. This ensures
that Cyc_end_thread is called with the correct thread data by the thread
which will be terminated.
This changes the behaviour to match r7rs (round x) instead of C round(x).
An answer to https://stackoverflow.com/questions/32746523/ieee-754-compliant-round-half-to-even
suggests using remainder(). The following will work if FE_TONEAREST is defined, but C11
requires FE_TONEAREST to be defined if and only if the implemenetation supports it in
fegetround() and fesetround() [Draft N1570]. On the other hand, remainder() must be defined.
C23 will have roundeven(), but this is not yet available on all platforms.
The behaviour of remainder is described in Draft N1570, page 254, footnote 239.
Alternative implementation:
double round_to_nearest_even(double x)
{
#pragma STDC FENV_ACCESS ON
int mode;
double nearest;
mode = fegetround();
fesetround(FE_TONEAREST);
nearest = nearbyint(x);
fesetround(mode);
#pragma STDC FENV_ACCESS OFF
return nearest;
}
Enforce a maximum C recursion depth when printing data structures. This protects against cases where a circular data structure may produce infinite output, blowing the stack. The recursive limit is sufficiently large such that a non-circular structure should not be impacted.
To fixes:
- Prevent segfault setting a global variable to itself
- Do not throw an error when exporting a primitive that is not defined in the current module, as built-ins are always available in any context.
