[Bug tree-optimization/19828] [4.0 Regression] LIM is pulling out a pure function even though there is something which can modify global memory

["joseph at codesourcery dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From joseph at codesourcery dot com  2005-02-13 20:38 -------
Subject: Re:  [4.0 Regression] LIM is pulling
 out a pure function even though there is something which can modify global
 memory

On Sun, 13 Feb 2005, drow at gcc dot gnu dot org wrote:

> I've suggested when talking to someone else about this that they should be
> assumed not to trap at the points where they are called.  That would allow
> calls to them to be removed, although still limit code motion.
> 
> It's a pity there's no rigorous definition of the current meaning of pure.

I would say that a pure function is a pure but not necessarily total 
function of its arguments and those parts of the state of the machine 
visible to it: you can move or remove calls to pure functions as long as 
in the abstract machine the function would have been called with the 
particular machine state with which it does get called.  So strlen is pure 
and "if (s) strlen(s);" is valid and the strlen call can't be moved before 
the test for NULL.  Similarly a const function is a pure but not 
necessarily total function of the values of its arguments only: so a 
square root function which doesn't set errno or floating point exception 
flags could be "const" and a call to it should not be moved before a test 
that the argument is nonnegative.

Where there are optimizations depending on the function being total - not 
trapping even if you call it with arguments with which the program never 
calls it in the abstract machine model - perhaps an additional attribute 
"total" should be added which can be used in conjunction with "const" or 
"pure".  (Though I'm not sure how many real functions are going to be 
"pure" and "total" but not "const".)



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19828