[Bug rtl-optimization/32296] [4.3 Regression] Bootstrap failure in stage1 on hppa*-*-*

["dave at hiauly1 dot hia dot nrc dot ca" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #7 from dave at hiauly1 dot hia dot nrc dot ca  2007-06-12 23:30 -------
Subject: Re:  [4.3 Regression] Bootstrap failure in stage1 on hppa*-*-*

> hppa_can_use_return_insn_p() is called 
> from "return" insn pattern in pa.md.
> The pattern looks:
> 
> (define_insn "return"
>   [(return)
>    (use (reg:SI 2))
>    (const_int 0)]
>   "hppa_can_use_return_insn_p ()"
>   "*
> 
> i.e. there's a use of reg 2.
> 
> Before dataflow branch got merged, regs_ever_live(2) was not true
> for reg 2, even though it is used at the return (hence it's live
> throughout the function body) - unless there's some other use
> within the function.
> 
> Now, with df, we always have the correct and precise regs_ever_live.
> Since the condition hppa_can_use_return_p() is checking 
> is really that there are no other insn that uses reg 2
> AND since we know the pattern has a use of reg 2,
> checking the number of use count to be 1
> checks effectively the same assertion.

Thanks for the detailed explanation.  I don't think the change
is correct.  The problem is the epilogue expander also calls
hppa_can_use_return_p().  It's not going to be true before before
the simple return is emitted.  As a result, we will never emit
a trivial return.

I think the solution is to remove this check from the insn definition
in pa.md and change the check in hppa_can_use_return_insn_p()
to ! df_hard_reg_used_p(). 

Dave


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32296