This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: input address reload issue


On 06/01/2017 17:06, Jeff Law wrote:
> On 01/06/2017 03:20 AM, Aurelien Buhrig wrote:
>>
>>>> So the insn:
>>>> (set (reg:QI 0 r0) (mem:QI (plus:SI (reg:SI 2 r2)(const_int 1))
>>>>
>>>> is transformed into:
>>>> (set (reg:SI 8 a0) (reg:SI 2 r2))
>>>> (set (reg:SI 8 a0) (const_int 1))
>>>> (set (reg:SI 8 a0) (plus:SI (reg:SI 8 a0) (reg:SI 8 a0)))
>>>> (set (reg:QI 0 r0) (mem:QI (reg:SI 8 a0))
>>>>
>>>> This "basic" transformation requires two reload regs, but only one is
>>>> given/used/possible from the rl structure (in emit_reload_insns).
>>>>
>>>> So where does the issue comes from? The need for 2 reload regs, the
>>>> transformation which is too "basic" and could be optimized to use only
>>>> one reload reg, or any wrong/missing reload target hook?
>>> Sounds like you need secondary or intermediate reloads.
>> Do you suggest the secondary reload must implement a scratch reg & md
>> pattern to implement this reload?
> Perhaps.  I don't know enough about your architecture to be 100% sure
> about how all the pieces interact with each other -- reload, and
> secondary reloads in particular are a complex area.  I'm largely going
> on your comment that you need 2 reload registers.
>
> Presumably you don't have an instruction for
> (set (reg) (plus (reg) (const_int)))
>
> Thus you need two scratch reload regs IIUC.  One to hold r2 another to
> hold (const_int 1).  So you'd want to generate
>
> (set (areg1) (reg r2))
> (set (areg2) (const_int 1))
> (set (areg1) (plus (areg1) (areg2)
> (set (r0) (mem (areg1))
>
> Or something along those lines.  If you're going to stick with reload,
> you'll likely want to dig into find_reloads_address and its children
> to see what reloads it generates and why (debug_reload can be helpful
> here). 

Thank you very much Jeff for your help. The best mapping for my target
would be:
(set areg1 r2) (set r0 (mem:QI (areg1 + 1)), and only 1 scratch would be
needed.

In secondary_reload, I can see the pattern (mem:QI (reg+1)) during IRA
pass with reg being a pseudo (ira_cost), but I cannot tranform it since
I do not know in which class the address reg would be reloaded into (or
maybe I can, but I don't know how since reg_renumber is -1 for this pseudo).
And the reload pass does not call my secondary reload with this pattern
mapped onto hw regs, so I cannot transform it.

So I try to add LEGITIMIZE_RELOAD_ADDRESS. I can see the (mem:QI (r2+1))
with hwreg, and so try to push_reload r2 with the "BASE_REG_CLASS".
But it has no effect. Indeed, in push_secondary_reload, the
secondary_reload is called and if NO_REG and CODE_FOR_nothing is
returned, "no reload reg is needed" ... But r2 is a hwreg which does not
belong to BASE_REG_CLASS, can be directly copied to BASE_REG_CLASS and
it does need a reload reg.
So should not push_secondary_reload check for class intersection ? Or
how to perform a reload from a hwreg to a class it does not belong to?

Thanks,
Aurélien





Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]