Re: ppc: weird movdi_internal32 interaction with FLOAT_REGS

[Geoff Keating <geoffk at geoffk dot org>]

On 09/08/2004, at 4:59 AM, Aldy Hernandez wrote:

> Hi Geoff.
>
> Could you or anyone else give me a hint as to why powerpc-eabispe is
> not building after your patch of (See below for ICE while building
> libstdc++):
>
> http://gcc.gnu.org/ml/gcc-patches/2004-08/msg00022.html
>
> I don't understand why this patch would cause an interaction with
> FLOAT_REGS, though I do see more 'f's before 'm's now :-).
>
> Perhaps I need a separate movdi_internal32 for SPE without the 'f'
> constraints?

It looks like reload decided to reload this insn by using the 'f -> m' 
alternative, not the 'r -> o<>' alternative.  I think that's probably 
bad, since if reload is being involved at all then r->o<> is going to 
be cheaper (if an FPR could have been used all the time, global-alloc 
should have picked it, and reloading the memory address will be cheaper 
than a GPR->FPR move).  Could you see why reload is doing this?  Maybe 
it's confused, or maybe something (like PREFERRED_RELOAD_CLASS) 
confused it.

> Thanks.
> Aldy
>
> a.ii: In function oid __static_initialization_and_destruction_0(int, 
> int)':
> a.ii:5631: error: unable to find a register to spill in class 
> `FLOAT_REGS'
> a.ii:5631: error: this is the insn:
> (insn 30 27 31 1 (set (mem/i:DI (lo_sum:SI (reg/f:SI 9 9 [139])
>                 (symbol_ref/i:SI 
> ("_ZGVN9__gnu_cxx10__mt_allocIcE10_S_optionsE") [flags 0x4] <var_decl 
> 0xf6c38174 _ZGVN9__gnu_cxx10__mt_allocIcE10_S_optionsE>)) [0 
> _ZGVN9__gnu_cxx10__mt_allocIcE10_S_optionsE+0 S8 A64])
>         (reg:DI 11 11 [orig:121 T.319 ] [121])) 328 
> {*movdi_internal32} (insn_list 27 (nil))
>     (expr_list:REG_DEAD (reg/f:SI 9 9 [139])
>         (nil)))
> a.ii:5631: internal compiler error: in spill_failure, at reload1.c:1884

Attachment: smime.p7s
Description: S/MIME cryptographic signature