Code bloat due to silly IRA cost model?

[Richard Sandiford]

Georg-Johann Lay <gjl@gcc.gnu.org> writes:
> Am 11.12.19 um 18:55 schrieb Richard Sandiford:
>> Georg-Johann Lay <gjl@gcc.gnu.org> writes:
>>> Hi, doesn't actually anybody know know to make memory more expensive
>>> than registers when it comes to allocating registers?
>>>
>>> Whatever I am trying for TARGET_MEMORY_MOVE_COST and
>>> TARGET_REGISTER_MOVE_COST, ira-costs.c always makes registers more
>>> expensive than mem and therefore allocates values to stack slots instead
>>> of keeping them in registers.
>>>
>>> Test case (for avr) is as simple as it gets:
>>>
>>> float func (float);
>>>
>>> float call (float f)
>>> {
>>>       return func (f);
>>> }
>>>
>>> What am I missing?
>>>
>>> Johann
>>>
>>>
>>> Georg-Johann Lay schrieb:
>>>> Hi,
>>>>
>>>> I am trying to track down a code bloat issue and am stuck because I do
>>>> not understand IRA's cost model.
>>>>
>>>> The test case is as simple as it gets:
>>>>
>>>> float func (float);
>>>>
>>>> float call (float f)
>>>> {
>>>>      return func (f);
>>>> }
>>>>
>>>> IRA dump shows the following insns:
>>>>
>>>>
>>>> (insn 14 4 2 2 (set (reg:SF 44)
>>>>          (reg:SF 22 r22 [ f ])) "bloat.c":4:1 85 {*movsf}
>>>>       (expr_list:REG_DEAD (reg:SF 22 r22 [ f ])
>>>>          (nil)))
>>>> (insn 2 14 3 2 (set (reg/v:SF 43 [ f ])
>>>>          (reg:SF 44)) "bloat.c":4:1 85 {*movsf}
>>>>       (expr_list:REG_DEAD (reg:SF 44)
>>>>          (nil)))
>>>> (note 3 2 6 2 NOTE_INSN_FUNCTION_BEG)
>>>> (insn 6 3 7 2 (set (reg:SF 22 r22)
>>>>          (reg/v:SF 43 [ f ])) "bloat.c":5:12 85 {*movsf}
>>>>       (expr_list:REG_DEAD (reg/v:SF 43 [ f ])
>>>>          (nil)))
>>>> (call_insn/j 7 6 8 2 (parallel [
>>>>
>>>> #14 sets pseudo 44 from arg register R22.
>>>> #2 moves it to pseudo 43
>>>> #6 moves it to R22 as it prepares for call_insn #7.
>>>>
>>>> There are 2 allocnos and cost:
>>>>
>>>> Pass 0 for finding pseudo/allocno costs
>>>>
>>>>      a1 (r44,l0) best NO_REGS, allocno NO_REGS
>>>>      a0 (r43,l0) best NO_REGS, allocno NO_REGS
>>>>
>>>>    a0(r43,l0) costs: ADDW_REGS:32000 SIMPLE_LD_REGS:32000 LD_REGS:32000
>>>> NO_LD_REGS:32000 GENERAL_REGS:32000 MEM:9000
>>>>    a1(r44,l0) costs: ADDW_REGS:32000 SIMPLE_LD_REGS:32000 LD_REGS:32000
>>>> NO_LD_REGS:32000 GENERAL_REGS:32000 MEM:9000
>>>>
>>>> which is quite odd because MEM is way more expensive here than any REG.
>>>>
>>>> Okay, so let's boost the MEM cost (TARGET_MEMORY_MOVE_COST) by a factor
>>>> of 100:
>>>>
>>>>      a1 (r44,l0) best NO_REGS, allocno NO_REGS
>>>>      a0 (r43,l0) best NO_REGS, allocno NO_REGS
>>>>
>>>>    a0(r43,l0) costs: ADDW_REGS:3200000 SIMPLE_LD_REGS:3200000
>>>> LD_REGS:3200000 NO_LD_REGS:3200000 GENERAL_REGS:3200000 MEM:801000
>>>>    a1(r44,l0) costs: ADDW_REGS:3200000 SIMPLE_LD_REGS:3200000
>>>> LD_REGS:3200000 NO_LD_REGS:3200000 GENERAL_REGS:3200000 MEM:801000
>>>>
>>>> What??? The REG costs are 100 times higher, and stille higher that the
>>>> MEM costs.  What the heck is going on?
>>>>
>>>> Setting TARGET_REGISTER_MOVE_COST and also TARGET_MEMORY_MOVE_COST to 0
>>>> yiels:
>>>>
>>>>    a0(r43,l0) costs: ADDW_REGS:0 SIMPLE_LD_REGS:0 LD_REGS:0 NO_LD_REGS:0
>>>> GENERAL_REGS:0 MEM:0
>>>>    a1(r44,l0) costs: ADDW_REGS:0 SIMPLE_LD_REGS:0 LD_REGS:0 NO_LD_REGS:0
>>>> GENERAL_REGS:0 MEM:0
>>>>
>>>> as expected, i.e. there is no other hidden source of costs considered by
>>>> IRA.  And even TARGET_REGISTER_MOVE_COST = 0  and
>>>> TARGET_MEMORY_MOVE_COST = original gives:
>>>>
>>>>    a0(r43,l0) costs: ADDW_REGS:32000 SIMPLE_LD_REGS:32000 LD_REGS:32000
>>>> NO_LD_REGS:32000 GENERAL_REGS:32000 MEM:9000
>>>>    a1(r44,l0) costs: ADDW_REGS:32000 SIMPLE_LD_REGS:32000 LD_REGS:32000
>>>> NO_LD_REGS:32000 GENERAL_REGS:32000 MEM:9000
>>>>
>>>> How the heck do I tell ira-costs that registers are way cheaper than MEM?
>> 
>> I think this is coming from:
>> 
>>    /* FIXME: Ideally, the following test is not needed.
>>          However, it turned out that it can reduce the number
>>          of spill fails.  AVR and it's poor endowment with
>>          address registers is extreme stress test for reload.  */
>> 
>>    if (GET_MODE_SIZE (mode) >= 4
>>        && regno >= REG_X)
>>      return false;
>> 
>> in avr_hard_regno_mode_ok.  This forbids SFmode in r26+ and means that
>> moves between pointer registers and general registers have the highest
>> possible cost (65535) to prevent them for being used for SFmode.  So:
>> 
>>     ira_register_move_cost[SFmode][POINTER_REGS][GENERAL_REGS] = 65535;
>> 
>> The costs for union classes are the maximum (worst-case) cost of
>> for each subclass, so this means that:
>> 
>>     ira_register_move_cost[SFmode][GENERAL_REGS][GENERAL_REGS] = 65535;
>> 
>> as well.
>> 
>> Removing the code above fixes it.  If you don't want to do that, an
>> alternative might be to add a class for r0-r25 (but I've not tested that).
>
> I am still having some headache understanding that...
>
> For example, currently R26 is forbidden for SFmode, but the same applies 
> to R25 or any odd registers (modes >= 2 regs have to start in even 
> registers).
>
> Then this would imply, even after the condition regno >= 26 was removed, 
> the costs would still be astronomically high because HI:21 is refused 
> and SI:23 is refused etc, and due to that the cost of that class will be 
> 0x10000 for modes >= 2 regs?

No, that will be OK.  The above happens at the class level, not at the
level of individual registers.  All classes that contain HI:21 (21-22)
or SI:23 (23-26) also contain valid HI and SI registers, so the costs
will be based on the valid registers.

The problem in the case above is that POINTER_REGS is big enough to hold
SFmode but isn't allowed to do so.

> How can the register allocator tell apart whether a register is rejected 
> due to its mode or due to the register number?  AFAIK there is no other 
> ws than rejecting odd registers in that hook, because register classes 
> must not have holes.  Or did that change meanwhile?

This hook is still the right way of rejecting odd registers.  The RA
doesn't really need to know whether something was rejected because of
its register number, mode, or both.

Thanks,
Richard