Richard, this sounds very good. Let me try this. So far, what I meant
by 'reloading myself' was that when generating the RTL above, I
actually emit move instructions to/from memory if the arguments are
memory, and generate RTL above only with hard registers. I did not
mention that RTL expansion is done from a builtin function (which is
where I am getting either memory or register arguments)..
Your approach sounds better.
On Tue, Sep 17, 2013 at 10:33 AM, Richard Sandiford
<rdsandif...@googlemail.com> wrote:
> Hendrik Greving <hendrik.greving.in...@gmail.com> writes:
>> For a special mechanism we are generating jump_insn with a 'set' side
>> effect in our backend. RTL looks e.g. like this:
>>
>> (jump_insn 25 24 26 (nil) (parallel [
>> (set (pc)
>> (if_then_else (ne (unspec_volatile [
>> (const_string ("<myinsn> %0,[%1] =%2"))
>> (const_int 0 [0x0])
>> (reg:SI 348)
>> ] 21)
>> (const_int 0 [0x0]))
>> (label_ref:SI 43)
>> (pc)))
>> (clobber (reg:SI 321 link))
>> (set (reg/v:QI 346)
>> (unspec:QI [
>> (const_int 0 [0x0])
>> ] 0))
>> ]) -1 (nil)
>> (nil))
>>
>> After working out some issues initially, this all works fine and the
>> 'set' seems to be properly recognized by RTL optimization phases (e.g.
>> CSE). I am now running into an issue however with reload. The problem
>> seems to be that if the parallel 'set' from e.g. RTL above feeds into
>> a (mem (reg)). This happens when e.g. compiling in debug mode, most
>> variables are at memory locations of the stack. In this case, compiler
>> needs to reload the instruction above (I am actually not sure why, but
>> I guess this could happen all the time). The further problem seems to
>> be that there is a hard constraint in reload that jump_insn can't have
>> output operands / reloads.
>>
>> if (GET_CODE (insn) == JUMP_INSN || GET_CODE (insn) == CALL_INSN)
>> no_output_reloads = 1;
>>
>> so what I am doing is, when generating RTL above, I always 'reload'
>> myself in the backend, generating mov's to/from memory, and always
>> making sure only hard registers are put into RTL above. The problem is
>> that I need hard registers for that. I would strongly prefer using
>> pseudo's, but pseudo's also seem to require reload.
>
> This might be what you mean by doing reload yourself, but the usual
> way of handling this is to add a memory alternative to the pattern
> and split that alternative after reload. If the split requires
> a temporary register, you can allocate one by adding a
> (clobber (match_scratch ...)) to the insn pattern. The scratch can
> be "X" for the normal register case that doesn't need a temporary.
> See the *ctr<mode>_internal1 pattern in config/rs6000/rs6000.md and
> the doloop_si64 pattern in config/s390/s390.md for examples.
>
> Thanks,
> Richard
