An issue for the SC: horrible documentation quality of GCC

[Jan Hubicka]

>     Well, when dead code is removed this will look like:
> 
>       p2 <-- 100 + p3
> 
> Only if p1 is, in fact, dead.
> 
>     But this needs to be determined at a later point, instead of cluttering
>     all the optimization passes with machine dependend checks.
> 
> You lost me here.  All of RTL and every choice within it is implicitly
> machine-dependent.  Every cost calculation is machine-dependent.  The
> entire RTL optimization strategy is based on making machine-dependent
> decisions (of course, written in machine-independent ways).
> 
>     CSE as we implement it is fundamentally different from gcse.  For one
>     our CSE is not any CSE but a kind of ad hoc value numbering
>     intertwined with expression simplification.
> 
> Again, you lost me.  It eliminates common subexpressions.  Isn't that what
> "cse" stands for?
No, the CSE algorithm as usually defined in the books would simplify
for instance:
(set (reg a) (plus b c))
(use a)
(set (reg a) (something else))
(set (reg b) (plus b c))

But our alrgorithm won't, but it would do different things CSE described
in the books won't, so it is not CSE as rest of the world know it and
naming it CSE is missleading.

Honza
> 
>     And I can't find a function in cse.c which does something similar
>     (i.e. just propagating constants and copies).  
> 
> Well, fold_rtx, I suppose, but that's because it does everything at
> one time.  And, as I've pointed out, it's not always better to
> substitute a constant for a register: sometimes, it's best to do it
> the other way around.  I have a favorite example for this:  the (non-obsolete)
> ROMP chip had an ABI where the return register and the first argument 
> register were the same. It also had a conditional return insn.  So if you
> have a function with one operand, say X, and you have:
> 
> 	if (x == 0)
> 	  return 0;
> 
> What you really want to do with the insn setting the return value is
> to change *from* the 0 *to* the equivalent, X.  That way that insn is a
> no-op, and you can do this as a conditional return insn.
> 
> CSE gets this right because it uses the cost metric.  It sounds like the
> code in gcse.c will get this wrong.
> 
>     Like I said, local_cprop* has nothing to do with gcse itself, except being
>     run to clean up the code between the gcse iterations.
> 
> Then that's all the *more* reason why it should be documented!