[fpc-devel] Pure function Wiki page

[J. Gareth Moreton]

 I've only really researched and improved the peephole optimizer, which is
the assembler stage.  I'm not sure how much optimisation is done at
earlier stages, but I do know that care is taken in deciding the best
implementation of a case block, for example, evaluating factors such as how
many separate branches there are.
 As it currently stands, pure function evaluation would be at the
pre-compilation stage, where Pascal code is converted into
platform-independent nodes.
 Gareth 

 On Mon 09/07/18 03:33 , "R0b0t1" r030t1 at gmail.com sent:
 n Sun, Jul 8, 2018 at 9:07 PM, Dmitry Boyarintsev 
  wrote: 
 > 
 > On Sun, Jul 8, 2018 at 8:15 PM, J. Gareth Moreton 
 >  wrote: 
 >> 
 >> Yes, if any parameters are variables, then the function is not
evaluated. 
 >> My intention is that the purity of a function is only determined when
it 
 >> comes to evaluating it in an expression, but because of how complex 
 >> functions can become, the "pure" directive hints to the compiler that
the 
 >> given function is pure and it should attempt the laborous task of
evaluating 
 >> it, rather than the opposite approach of attempting to evaluate all 
 >> functions with constant actual parameters and potentially increasing
the 
 >> compilation time by several orders of magnitude (don't forget it might
be 
 >> attempting to do the same thing with system functions if the project is

 >> undergoing a full build). 
 > 
 > 
 > if FPC assembler reader powerful enough to analyze and trust assembler 
 > functions marked as pure? 
 > 

 At which stages is optimization done? GCC's backend optimizes each 
 step in compilation (as I am aware), i.e. GENERIC -> GIMPLE -> RTL -> 
 assembly. Many optimizations work best or are only possible at a 
 certain stage. 

 The various representations are also what makes analysis efficient. 
 Most optimization passes do not happen on assembly. 

 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freepascal.org/pipermail/fpc-devel/attachments/20180709/98c228c7/attachment.html>