The previous stackless implementations marked all 'luaV_execute'
invocations as fresh. However, re-entering 'luaV_execute' when
resuming a coroutine should not be a fresh invocation. (It works
because 'unroll' called 'luaV_execute' for each call entry, but
it was slower than letting 'luaV_execute' finish all non-fresh
invocations.)
A "with stack" implementation gains too little in performance to be
worth all the noise from C-stack overflows.
This commit is almost a sketch, to test performance. There are several
pending stuff:
- review control of C-stack overflow and error messages;
- what to do with setcstacklimit;
- review comments;
- review unroll of Lua calls.
When "undumping" a long string, the function 'loadVector' can call the
reader function, which can run the garbage collector, which can collect
the string being read. So, the string must be anchored during the call
to 'loadVector'.
Avoid turning an object to gray except at the moment it is inserted in a
gray list or in the explicit exceptional cases such as open upvalues and
fixed strings.
- Macro 'gray2black' was renamed 'nw2black' (Non-White to black), as it
was already being used on objects that could be already black.
- Macros 'white2gray' and 'black2gray' were unified in 'set2gray'; no
reason to have two macros when one will do and, again, 'black2gray' was
already being used on objects that could be already gray.
Moreover, macros 'maskcolors' and 'maskgcbits' were negated to have
ones in the described bits, instead of zeros. (This naming seems more
intuitive.)
This commit fixes a bug introduced in commit 9cf3299fa. TOUCHED2
objects are always black while the mutator runs, but they can become
temporarily gray inside a minor collection (e.g., if the object is a
weak table).
Tables were using this bit to indicate their array sizes were real
('isrealasize'), but this bit can be useful for tests. Instead, they
can use bit 7 of their 'flag' field for that purpose. (There are only
six fast-access metamethods.) This 'flag' field only exists in tables,
so this use does not affect other types.
In incremental mode, threads don't need to be visited again once
visited in the atomic phase. In generational mode (where all visits
are in the atomic phase), only old threads need to be kept in the
'grayagain' list for the next cycle.
When entering generational mode, all objects are old. So, the only
objects that need to be in a gray list are threads, which can be
assigned without barriers. Changes in anything else (e.g., weak
tables) will trigger barriers that, if needed, will add the object
to a gray list.
OLD1 objects can be potentially anywhere in the 'allgc' list (up
to 'reallyold'), but frequently they are all after 'old1' (natural
evolution of survivals) or do not exist at all (when all objects die
young). So, instead of 'markold' starts looking for them always
from the start of 'allgc', the collector keeps an extra pointer,
'firstold1', that points to the first OLD1 object in the 'allgc' list,
or is NULL if there are no OLD1 objects in that list.
Instead of adding all tables and userdata back to the 'grayagain' list
to be checked by 'correctgraylist', the collector adds only the objects
that will remain in that list (objects aged TOUCHED1). This commit
also rewrites 'correctgraylist' with a clearer logic.
Small changes to ensure that all objects are kept 'new' in incremental
GC (except for fixed strings, which are always old) and to make that
fact clearer.
Barriers cannot be active during sweep, even in generational mode.
(Although gen. mode is not incremental, it can hit a barrier when
deleting a thread and closing its upvalues.) The colors of objects are
being changed during sweep and, therefore, cannot be trusted.
In luaD_callnoyield, when there is a possible stack overflow, it
zeros the number of CallInfos to force a check when calling the
function. However, if the "function" is not a function, the code will
raise an error before checking the stack. Then, the error handling calls
luaD_callnoyield again and nCcalls is decremented again, crossing the
stack redzone without raising an error. (This loop can only happens
once, because the error handler must be a function. But once is enough
to cross the redzone.)
The field 'L->oldpc' is not always updated when control returns to a
function; an invalid value can seg. fault when computing 'changedline'.
(One example is an error in a finalizer; control can return to
'luaV_execute' without executing 'luaD_poscall'.) Instead of trying to
fix all possible corner cases, it seems safer to be resilient to invalid
values for 'oldpc'. Valid but wrong values at most cause an extra call
to a line hook.
When entering a coroutine, the computation of nCcalls added 'from->nci'
to correct for preallocated CallInfos, but 'nci' includes also the
Callinfos already used.
When an object aged OLD1 is finalized, it is moved from the list
'finobj' to the *beginning* of the list 'allgc'. So, this part of the
list (and not only the survival list) must be visited by 'markold'.
Macro HARDMEMTESTS broke in two: HARDMEMTESTS forces a full GC cycle
at every point where the GC can run. New macro EMERGENCYGCTESTS
forces an emergency collection at every memory allocation.
Errors in finalizers need a valid 'pc' to produce an error message,
even if the error is not propagated. Therefore, calls to the GC (which
may call finalizers) inside luaV_execute must save the 'pc'.
Macro 'checkstackGC' was doing a GC step after resizing the stack;
the GC could shrink the stack and undo the resize. Moreover, macro
'checkstackp' also does a GC step, which could remove the preallocated
CallInfo when calling a function. (Its name has been changed to
'checkstackGCp' to emphasize that it calls the GC.)
Allow memory errors to be raised through the API (throwing the
error with the memory error message); error in external allocations
raises a memory error; memory errors in coroutines are re-raised
as memory errors.
To allow their use in memory tests, some functions in 'ltests.c'
should never allocate memory. To avoid this allocation, the
library registers the strings used for status codes, and keeps
the variable '_WARN' always defined (with false instead of nil).
