Re: Bugs in SSA-CCP and SSA-PRE optimizers

[Daniel Berlin <dberlin at dberlin dot org>]

On Tuesday, September 17, 2002, at 03:58  PM, Steven Bosscher wrote:

> Hello,
>
> Consider:
>
> int main()
> {
>   int bla, foo, bar;
>
>   bla = 1;
>   foo = 2*bla;
>   bar = 2*bla+1;
>   goto waffle;
>   return bar;
> waffle:
>   bla = 1;
>   return bar;
> }
>
> Obviously quite simple to do CCP and PRE on (and DCE but that's not in
> CVS yet).
>
> So I tried to compile it and see what the optimized tree looks like. I
> applied a patch to dump the optimized tree after the SSA optimizations
> are done.
>
> Compiler is this mornings tree-ssa-branch.
> Left hand side is: gcc -fdump-tree-all-ssa t.c
> Right hand side is:
> gcc -fdump-tree-all-ssa -ftree-ssa -ftree-ssa-pre -ftree-ssa-ccp t.c
Stop.
already, you are screwed.
CCP doesn't update the SSA representation, thus, PRE won't have correct 
data.

Try them individually, not together, for right now.

> diff -y 1/t.c.t10.optimized 3/t.c.t10.optimized
> main()					main()
> {					{
>   int bla;				  int bla;
>   int foo;				  int foo;
>   int bar;				  int bar;
>   int T.1;				  int T.1;
> 					> int pretmp.2;
>   bla = 1;				  bla = 1;
>   foo = bla * 2;		      |	  foo = pretmp.2 = 1 * 2;
>   T.1 = bla * 2;		      |	  T.1 = pretmp.2;
>   bar = T.1 + 1;			  bar = T.1 + 1;
>   goto waffle;				  goto waffle;
>   return bar;				  return bar;
>   waffle:				  waffle:
>   bla = 1;				  bla = 1;
>    return bar;				  return bar;
> }					}
>
> Note the redundant assignment to T.1. It seems PRE doesn't kill those
> T.[0-9*] variables (both uses and it declarations). We're generating (a
> lot of?) useless RTL here.
> Also, why is "foo = pretmp.2 = 1 * 2" not folded?
Because i don't call fold on the result?