Re: advanced value range propagation for tree-ssa

[Jeff Law <law at redhat dot com>]

zhouyi zhou wrote:
> High, 
>    Currently gcc can optimize following program by threading a block:
> int
> fn(int s, int w) {
>   int i;
>
>   i = 2;
>
>   if (s)
>    i = 1;
>
>   if (s)
>    printf("%d\n", i);
>
>   return 0;
> }
>
>   While it cannot handle programs of following kind:
> int
> fn(int s, int w) {
>   int i;
>
>   i = 2;
>
>   if (s)
>    i = 1;
>
>   if (w)
>    goto out1;
>
>   if (s)
>    printf("%d\n", i);
>
> out1:
>   return 0;
> }
>
> The advanced value range propagation:
> http://wiki.freebsd.org/ZhouyiZHOU?action=AttachFile&do=get&target=avrp.patch
>
> nicely handles the above case
>
>       2 (if s)
>     /   \
>    8     3
>     \     |
>      \__ 4----
>         / 
>       5(if s)  
>     /  \
>    6   10
>
> The advanced value range propagation 
> 1) identify the candidate basic block pairs by 
> recurively matching the dom son and dom parent (in our case block 5 and block 2 are
> selected)
> 2) identiy the threadable edge pairs:
> here edge pair 8-4 and 5-6, 3-4 and 5-10 are selected
> 3) thread the edges (duplicate the basic blocks between)
>       2
>     /  \
>    8    3
>    |     \
>    4'-    4''-
>    |      |
>    5'     5''
>   /       |
>  6        10
>
> The patch is being compiled and bootstraped on i686-linux m-32.
> and passed most of regression test by make -k check (the not passed tests
>  are caused the wrong configuration of my machine like libgmp and libmpfr)
>
> The patch is rudimentary and very dirty written currently, I will clean it in the future
Kenny outlined an algorithm which should handle these cases rather 
easily at the GCC summit a few years back.   You might ask him for the 
slides from his opening presentation as they included the outline of the 
algorithm.

I did an implementation to compare it against the current edge threading 
code and while it didn't stack up too well, there were a few cases where 
Kenny's algorithm would find something optimizable, but the current 
threader would not.  Your example is one of those cases.

Implementation was fairly easy -- it might be easier to work from 
Kenny's algorithm than cleaning up your own.


Jeff
>