question on TARGET_SECONDARY_RELOAD

question on TARGET_SECONDARY_RELOAD

[Henri Cloetens]

Hello all,

When writing my custom backend, I hit problems when more processor 
registers are needed than there
are available.

In this case,
   - it works without the optimizer (Obviously . Without the optimizer, 
the compiler puts all on the stack.)
   - it crashes with the optimizer. It crashes in the reload step, seems 
to me it fails to put some variables in memory.

This seems to be associated with my implementation of 
preferred_reload_class and secondary_reload.

This is from the log:

/blueproc_secondary_reload , return NO_REGS, rclass = BASE_REGS, mode = 
SI, type 1 x://
//(reg/v:SI 333 [ ndigits ])//
//
//blueproc_preferred_reload_class, return BASE_REGS, rclass = BASE_REGS, 
mode = SI, x://
//(reg/v:SI 333 [ ndigits ])//
//
//blueproc_preferred_reload_class, return BASE_REGS, rclass = BASE_REGS, 
mode = SI, x://
//(reg/v:SI 333 [ ndigits ])//
/

Now, something is obviously wrong. Secondary reload returns /no_regs/, 
meaning it should go straigth to memory.
But, question, what should I do with the /sri->code/ and /sri->cost/ 
arguments ??.

Best Regards,

Henri.

Re: question on TARGET_SECONDARY_RELOAD

[Segher Boessenkool]

Hi!

On Wed, Oct 21, 2020 at 11:40:46AM +0200, Henri Cloetens wrote:
> But, question, what should I do with the /sri->code/ and /sri->cost/ 
> arguments ??.

"icode" is the number of the RTL pattern that implements this.  "cost"
usually is the number of machine instructions you need to insert.


Segher