Re: Register Allocation issues with microblaze-elf

[Michael Eager <eager at eagerm dot com>]

On 02/13/2013 02:38 PM, Vladimir Makarov wrote:
> On 13-02-13 1:36 AM, Michael Eager wrote:
> > Hi --
> >
> > I'm seeing register allocation problems and code size increases
> > with gcc-4.6.2 (and gcc-head) compared with older (gcc-4.1.2).
> > Both are compiled using -O3.
> >
> > One test case that I have has a long series of nested if's
> > each with the same comparison and similar computation.
> >
> >         if (n<max_no){
> >           n+=*(cp-*p++);
> >           if (n<max_no){
> >             n+=*(cp-*p);
> >               if (n<max_no){
> >         . . .          ~20 levels of nesting
> >                <more computations with 'cp' and 'p'>
> >                 . . . }}}
> >
> > Gcc-4.6.2 generates many blocks like the following:
> >     lwi    r28,r1,68    -- load into dead reg
> >     lwi    r31,r1,140    -- load p from stack
> >     lbui    r28,r31,0
> >     rsubk    r31,r28,r19
> >     lbui    r31,r31,0
> >     addk    r29,r29,r31
> >     swi    r31,r1,308
> >     lwi    r31,r1,428    -- load of max_no from stack
> >     cmp    r28,r31,r29    -- n in r29
> >     bgeid    r28,$L46
> >
> > gcc-4.1.2 generates the following:
> >     lbui    r3,r26,3
> >     rsubk    r3,r3,r19
> >     lbui    r3,r3,0
> >     addk    r30,r30,r3
> >     swi    r3,r1,80
> >     cmp    r18,r9,r30    -- max_no in r9, n in r30
> >     bgei    r18,$L6
> >
> > gcc-4.6.2 (and gcc-head) load max_no from the stack in each block.
> > There also are extra loads into r28 (which is not used) and r31 at
> > the start of each block.  Only r28, r29, and r31 are used.
> >
> > I'm having a hard time telling what is happening or why.  The
> > IRA dump has this line:
> >    Ignoring reg 772, has equiv memory
> > where pseudo 772 is loaded with max_no early in the function.
> >
> > The reload dump has
> > Reloads for insn # 254
> > Reload 0: reload_in (SI) = (reg/v:SI 722 [ max_no ])
> >     GR_REGS, RELOAD_FOR_INPUT (opnum = 1)
> >     reload_in_reg: (reg/v:SI 722 [ max_no ])
> >     reload_reg_rtx: (reg:SI 31 r31)
> > and similar for each of the other insns using 722.
> >
> > This is followed by
> >   Spilling for insn 254.
> >   Using reg 31 for reload 0
> > for each insn using pseudo 722.
> >
> > Any idea what is going on?
> So many changes happened since then (7 years ago), that it is very hard to me to say something
> definitely.  I also have no gcc-4.1 microblaze (as I see microblaze was added to public gcc for 4.6
> version) and it makes me even more difficult to say something useful.
>
> First of all, the new RA was introduced in gcc4.4 (IRA) which uses different heuristics
> (Chaitin-Briggs graph coloring vs Chow's priority RA).
>
> We could blame IRA when we have the same started conditions for it RA gcc4.1 and gcc4.6-gcc-4.8.
> But I am sure it is not the same. More aggressive optimizations creates higher register pressure.  I
> compared peak reg pressure in the test for gcc4.6 and gcc4.8.  It became higher (from 102 to 106).
> I guess the increase was even bigger since gcc4.1.

I thought about register pressure causing this, but I think that should cause
spilling of one of the registers which were not used in this long sequence,
rather than causing a large number of additional loads.

Perhaps the cost analysis has a problem.

> RA focused on generation of faster code.  Looking at the fragment you provided it, it is hard to say
> something about it.  I tried -Os for gcc4.8 and it generates desirable code for the fragment in
> question (by the way the peak register pressure decreased to 66 in this case).

It's both larger and slower, since the additional loads take much longer.  I'll take a
look at -Os.

It looks like the values of p++ are being pre-calculated and stored on the stack.  This results in
a load, rather than an increment of a register.

> Any industrial RA uses heuristic algorithms, in some cases better heuristics can work worse than
> worse heuristics.  So you should probably check is there any progress moving from gcc4.1 to gcc4.6
> with performance point of view for variety benchmarks.  Introducing IRA improves code for x86 4% on
> SPEC2000.  Subsequent improving (like using dynamic register classes) made further performance
> improvements.

My impression is that the performance is worse.  Reports I've seen are that the code is
substantially larger, which means more instructions.

I'm skeptical about comparisons between x86 and RISC processors.  What works well
for one may not work well for the other.

> Looking at the test code, I can make some conclusions for myself:
>
> o We need a common pass decreasing reg pressure (I already expressed this in the past) as
> optimizations become more aggressive.  Some progress was made to make few optimizations aware about
> RA (reg-pressure scheduling, loop-invariant motions, and code hoisting) but there are too many
> passes and it is wrong and impossible to make them all aware of RA.  Some register pressure
> decreasing heuristics are difficult to implement in RA (like insn rearrangements or complex
> rematerialization) and this pass could focus on them.

That might be useful.

> o Implement RA live range splitting in regions different from loops or BB (now IRA makes splitting
> only on loop bounds and LRA in BB, the old RA had no live range splitting at all).

Each of the blocks of code is in it's own BB.  I haven't checked, but I'd guess
that most of the registers are in use on entry and still live on exit, so the
block has no registers to allocate.

> I'd also recommend to try the following options concerning RA: -fira-loop-pressure,
> -fsched-pressure, -fira-algorithm=CB|priority, -fira-region=one,all,mixed.  Actually
> -fira-algorithm=priority + -fira-region=one is analog of what the old RA did.

I'll see if that makes any difference.

> I hope I answered to your question.

Thanks.


--
Michael Eager	 eager@eagercon.com
1960 Park Blvd., Palo Alto, CA 94306  650-325-8077