Re: constant hoisting out of loops

[fanqifei <fanqifei at gmail dot com>]

On Mon, Mar 15, 2010 at 5:24 AM, Jim Wilson <wilson@codesourcery.com> wrote:
> On 03/10/2010 10:48 PM, fanqifei wrote:
>>
>> For below piece of code, the instruction "clr.w a15" obviously doesn't
>> belong to the inner loop.
>> ? ?6: ? bd f4 ? ? ? ? ? ?clr.w a15; #clear to zero
>> ? ?8: ? 80 af 00 ? ? ? ?std.w a10 0x0 a15;
>
> There is info lacking here. ?Did you compile with optimization? ?What does
> the RTL look like before and after the loop opt passes?
>
> I'd guess that your movsi pattern is defined wrong. ?You probably have
> predicates that allow either registers or constants in the set source, which
> is normal, and constraints that only allow registers when the dest is a mem.
> ?But constraints are only used by the reload pass, so a store zero to mem
> rtl insn will be generated early, and then fixed late during the reload
> pass. ?So the loop opt did not move the clear insn out of the loop because
> there was no clear insn at this time.
>
> The way to fix this is to add a condition to the movsi pattern that excludes
> this case. ?For instance, something like this:
> ? "(register_operand (operands[0], SImode)
> ? ? || register_operand (operands[1], SImode))"
> This will prevent a store zero to mem RTL insn from being accepted. ?In
> order to make this work, you need to make movsi an expander that accepts
> anything, and then forces the source to a register if you have a store
> constant to memory. ?See for instance the sparc_expand_move function or the
> mips_legitimize_move function.
>
> Use -da (old) or -fdump-rtl-all (new) to see the RTL dumps to see what is
> going on.
>
> Jim
>
It's compiled with -O2.
You are correct. The reload pass emitted the clr.w insn.
However, I can see loop opt passes after reload:
problem1.c.174r.loop2_invariant1
problem1.c.174r.redo_loop2_invariant
problem1.c.175r.loop2_unswitch
problem1.c.177r.redo_loop2_invariant
After reload pass, the clr.w insn is in the loop. And after above
loop2 passes, the insn is not moved outside of the loop.
I am not sure the issue is in these loop2 passes. I guess there is.

For the definition of movsi expander, I will try to do what you pointed out.
(I am not very familiar with these code and that may take me some time.)

current definition of mov pattern:
(define_insn "mov<mode>"
  [(set
    (match_operand:BWD 0 "nonimmediate_operand"      "=r,m,r,r,r,r,r,r,x,r")
    (match_operand:BWD 1 "move_source_operand"       "Z,r,L,I,Q,P,ni,x,r,r"))]
  ""
  "@
     %L1<m> %0 %1;
     %S0<m> %0 %1;
     clr<m> %0;
     mv     %0 %1;
   ... ...

Thanks!

-- 
-Qifei Fan
http://freshtime.org