[Bug rtl-optimization/64895] RA picks the wrong register for -fipa-ra

["vmakarov at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64895

Vladimir Makarov <vmakarov at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vmakarov at gcc dot gnu.org

--- Comment #7 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
(In reply to vries from comment #6)
> From PR64342 comment 7:
> 
> Register allocation seems to progress similarly, up until this message in
> reload, which seems to be directly related to the r216154 patch:
> ...
>     Spill r86 after risky transformations
>   Reassigning non-reload pseudos
>            Assign 3 to r86 (freq=3000)
> ...
> 
> I've looked a bit in more detail at this message.
> 
> fipa-ra manages to allocate hardreg 1 (dx) to the call-crossing liverange as
> before. But setup_live_pseudos_and_spill_after_risky_transforms looks at the
> conflict regs, and finds that lra_reg_info[regno].conflict_hard_regs
> contains dx. Hence, it spills the liverange into an available reg, which
> happens to be hardreg 3 (bx).
> 
> So, the risky transformations and fipa-ra have in common that they allocate
> registers from the conflict_hard_regs set. In the case of -fipa-ra, we don't
> want setup_live_pseudos_and_spill_after_risky_transforms to undo the
> -fipa-ra allocation, but currently there's no way to distinguish between the
> two.
> 

Right.  Introduction of pic pseudo and possible rematerializations of memory
references with the pic pseudo made checking the new conflict first necessary. 
Before this it was not a problem as pic hard reg was used only for pic
addressing.

> Using this patch, we make sure the conflict registers don't include dx for
> this testcase, and the test-cases passes again:
> ...
> diff --git a/gcc/lra-lives.c b/gcc/lra-lives.c
> index 9dfffb6..0231057 100644
> --- a/gcc/lra-lives.c
> +++ b/gcc/lra-lives.c
> @@ -636,8 +636,11 @@ check_pseudos_live_through_calls (int regno)
>    if (! sparseset_bit_p (pseudos_live_through_calls, regno))
>      return;
>    sparseset_clear_bit (pseudos_live_through_calls, regno);
> +
>    IOR_HARD_REG_SET (lra_reg_info[regno].conflict_hard_regs,
> -                   call_used_reg_set);
> +                   !hard_reg_set_empty_p
> (lra_reg_info[regno].actual_call_used_reg_set)
> +                   ? lra_reg_info[regno].actual_call_used_reg_set
> +                   : call_used_reg_set);
>  
>    for (hr = 0; hr < FIRST_PSEUDO_REGISTER; hr++)
>      if (HARD_REGNO_CALL_PART_CLOBBERED (hr, PSEUDO_REGNO_MODE (regno)))
> ...

The patch looks ok to me.  Tom, could you prepare the patch (check it mostly
for x86-64 bootstrap and testsuite) and commit it to the trunk.  I approve it.