Re: [RFA] The Integrated Register Allocator

[Vladimir Makarov <vmakarov at redhat dot com>]

Kaz Kojima wrote:
> Vladimir Makarov <vmakarov@redhat.com> wrote:
>   
> > After latest patches I have (on ira branch)
> >
> > _dadd:
> >     mov.l    r14,@-r15
> >     mov.l    .L3,r0
> >     sts.l    pr,@-r15
> >     jsr    @r0
> >     mov    r15,r14
> >     mov    r14,r15
> >     lds.l    @r15+,pr
> >     rts   
> >     mov.l    @r15+,r14
> >     
>
> This looks the code generated without -m4.
>  
>   
> > I think the patch responsible for fixing was
> >
> > http://gcc.gnu.org/ml/gcc-patches/2008-04/msg00400.html
> >     
>
> I've confirmed that the above patch already exists in my ira
> tree.  With -O2 -fira -m4, the insns in .174r.sched1 dump are
>
> (insn:HI 7 24 12 3 fadd.c:1 (parallel [
>             (set (reg:DF 162)
>                 (plus:DF (reg:DF 68 fr4 [ x ])
>                     (reg:DF 70 fr6 [ y ])))
>             (use (reg/v:PSI 151 ))
>         ]) 343 {adddf3_i} (expr_list:REG_DEAD (reg:DF 70 fr6 [ y ])
>         (expr_list:REG_DEAD (reg:DF 68 fr4 [ x ])
>             (nil))))
>
> (insn:HI 12 7 18 3 fadd.c:1 (parallel [
>             (set (reg/i:DF 64 fr0 [ <result> ])
>                 (reg:DF 162))
>             (use (reg/v:PSI 151 ))
>             (clobber (scratch:SI))
>         ]) 198 {movdf_i4} (expr_list:REG_DEAD (reg:DF 162)
>         (expr_list:REG_DEAD (reg/v:PSI 151 )
>             (nil))))
>
> and .178r.ira says
>
> Pass 0 for finding allocno costs
>
>     a0 (r162,l0) best GENERAL_REGS, cover GENERAL_REGS
>
>   a0(r162,l0) costs: SIBCALL_REGS:2000000 GENERAL_REGS:2000000 FP_REGS:2000000 MEM:2000000
>
>   
The problem was in rejecting all alternatives of insn 7 (because of 3rd 
operand r151 in more details) during register class cost calculation.
> though the old allocator choses r162 from FP_REGS.
>   
I've committed a patch fixing the problem.  Please see

http://gcc.gnu.org/ml/gcc-patches/2008-04/msg00516.html

After the patch I have the following code (when -O2 -fira -m4 are

used).

_dadd:
	fmov	fr4,fr0
	fmov	fr5,fr1
	mov.l	r14,@-r15
	mov	r15,r14
	fadd	dr6,dr0
	mov	r14,r15
	rts	
	mov.l	@r15+,r14