regarding df_regs_ever_live_p

[Jeff Law]

On 4/7/19 10:44 AM, William Tambe wrote:
> When STATIC_CHAIN_REGNUM is defined and the static chain register is
> not a fixed register, I am assuming that the prologue need to save it
> and that df_regs_ever_live_p must be checked to determine whether to
> save it.
Again, it depends on the ABI and what it says about the register.  Most
(if not all) the ABIs I've worked with through the years use a static
chain that is call clobbered and thus you don't have to do anything special.


> 
> The uncertainty arise when one cannot tell which GCC internal variable
> must be checked beside df_regs_ever_live_p ; ie:
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/fr30/fr30.c#L145
I'm not familiar with the fr30.  But AFAICT its static chain (%r12) is
clobbered.

In general it's better to include the actual code you're referencing.  I
don't have gcc-8.3.0 handy and the line number references are going to
be "off".

I'm guessing you're referring to this:

> /* Tell prologue and epilogue if register REGNO should be saved / restored.
>    The return address and frame pointer are treated separately.
>    Don't consider them here.  */
> #define MUST_SAVE_REGISTER(regno)      \
>   (   (regno) != RETURN_POINTER_REGNUM \
>    && (regno) != FRAME_POINTER_REGNUM  \
>    && df_regs_ever_live_p (regno)      \
>    && ! call_used_regs [regno]         )
> 
> #define MUST_SAVE_FRAME_POINTER  (df_regs_ever_live_p (FRAME_POINTER_REGNUM)  || frame_pointer_needed)
> #define MUST_SAVE_RETURN_POINTER (df_regs_ever_live_p (RETURN_POINTER_REGNUM) || crtl->profile)

Which looks pretty normal.  For a typical register you just need to know
if it's call saved (! call_used_regs) and if it's ever used
(df_regs_ever_live).

Frame pointers are sometimes handled specially (and the frame pointer is
different than the static chain).  Frame pointers are often eliminated
and references turned into a stack pointer address.  That in turn allows
the frame pointer to be used by the register allocator.

On the fr30 the return address appears to be handled specially as well.
 Not really sure why -- perhaps the call to mcount isn't fully exposed
or something similar.  It's not terribly uncommon to handle a few things
specially when profiling.


> 
> Hence the reason for asking for examples showing how to properly have
> the DF entry block define registers so that it is enough to check only
> df_regs_ever_live_p in the prologue on whether to save them.
> 
> Are they currently such examples ?
So the first question I would ask is whether or not your static chain is
a call clobbered register is not.  If not, why -- most ports have a call
clobbered static chain.

Jeff