Re: [avr-llvm-devel] Interrupt handling

[Borja F. <bor...@gm...>]

Ok lets go by parts:

I will fix the naked attribute to not emit the ret/reti instructions. I
don't know in gcc, but in llvm the ret instruction is inserted very early,
not during prologue/epilogue insertion, that's why it's getting emitted by
default. So additional code has to be added to handle this. Yes, I've tried
naked in non ISR functions aswell so I knew it was something valid to do. I
wasn't sure about how it worked in combination with other attributes.

About the interrupt and signal attributes, i think it should be possible to
error out if they're both defined. But looking in avr-libc's code when you
add the ISR_NOBLOCK macro inside the ISR macro you're both defining signal
and interrupt at the same time, so it would always error out, or am I
missing something?

#define ISR(vector, ...)            \
    void vector (void) __attribute__ ((signal,__INTR_ATTRS)) __VA_ARGS__; \
    void vector (void)
#  define ISR_NOBLOCK    __attribute__((interrupt))

2012/3/5 Weddington, Eric <Eri...@at...>

> Hi Borja,
>
> From my view, "interrupt" and "signal" are mutually exclusive. They both
> define an ISR, where "interrupt" allows nested interrupts and "signal"
> disallows nested interrupts.
>
> I don't know if we can error out if a user defines both simultaneously.
>
> "naked" does something completely different than interrupt|signal. One
> can have a naked function without it being an ISR.
>
> Can one have a "naked" ISR? I suppose if one is writing the ISR in
> assembly... Another reason to have interrupt|signal is to put that
> function into the IVT. So, yes, the combination is valid, but naked
> should have priority, and the prologue/epilogue should not be generated
> for that ISR, and it is up to the user to write.
>
> Technically, the RET/RETI is usually a part of the epilogue. Naked
> functions do not have prologues/epilogues, hence, they do not have a
> RET/RETI, and the user has to provide those.
>
> Hope that answers any questions. Again, my only suggestion would be to
> error out on simultaneous "interrupt" and "signal" attributes.
>
> Eric Weddington
>
> > -----Original Message-----
> > From: Borja Ferrer [mailto:bor...@gm...]
> > Sent: Monday, March 05, 2012 1:58 PM
> > To: avr...@li...
> > Subject: [avr-llvm-devel] Interrupt handling
> >
> > I'm currently implementing support for interrupt handling. But before
> going
> > any further into implementing it in the backend (the frontend part is
> nearly
> > done) I want to get things clear. Eric this one is probably for you
> because
> > it's about avr-libc but anyways, here we go:
> >
> > There are 2 function attributes to add:
> >
> > 1) signal (ISR macro): normal interrupt handler. Every register used
> inside
> > the handler gets saved in the stack, including special code to save
> SREG. So
> > we have to mark all registers to be callee saved.
> > 2) interrupt (ISR_NOBLOCK): same as above but adding a sei instruction
> at the
> > very top of function entry.
> >
> > Now, the naked attribute is already implemented and i don't need to
> handle
> > anything with it. It avoids any prologue/epilogue code to be inserted.
> BUT by
> > default the ret/reti instruction gets emitted, and avr-libc says that
> it
> > should be done by the programmer so I have to handle that, right?
> > This doesn't matter if it's an ISR or a normal func.
> >
> > The ISR_ALIASOF macro works correctly, so no more work is needed there
> (using
> > NOBLOCK or NAKED with ALIASOF doesn't make sense at all, right?)
> >
> > Now i want to talk about combining attributes:
> > Truth table format: |signal | interrupt | naked|
> >
> > 1) 0 0 0 -> nothing
> >
> > 2) 0 0 1 -> no prologue/epilogue and no ret/reti (used in any
> arbitrary
> > function)(so as i said before, don't emit ret/reti?)
> >
> > 3) 0 1 0 -> emit sei, and save all used regs inside the function plus
> SREG.
> >
> > 4) 0 1 1 -> naked wins, do case 1.
> >
> > 5) 1 0 0 -> normal ISR macro, do what i said when explaining the
> signal
> > attrib.
> >
> > 6) 1 0 1 -> naked wins again.
> >
> > 7) 1 1 0 -> interrupt wins, do case 3.
> >
> > 8) 1 1 1 -> naked wins.
> >
> > So this is what I understand I have to do, please let me know if I'm
> wrong on
> > anything I've mentioned.
> >
>
>