Re:[PING] Patch Fix PR32871 -- [avr] gcc is pushing too many registers

[Andy H <hutchinsonandy at aim dot com>]

Anatoly,

can you approve this so I can comit to avr mainline? It's been listed 
for a long time.

best regards

Andy


Andy H wrote:
> The attached patch fixs:
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32871
>
> Although the bug is against the AVR target it  reveals a weakness in 
> using df_regs_ever_live_p.
> to determine register to save and restore that may apply to other 
> targets.
>
> The fix determines which registers are used as arguments to function 
> using
> INIT_CUMULATIVE_ARGS, FUNCTION_ARG and FUNCTION_ARG_ADVANCE
> (This is patch is thus relatively portable for other targets)
>
> The argument registers  are then omitted from epilog/prolog save and 
> restores
> (as all argument registers are assumed clobbered.)
>
> Tested against HEAD and had no new regression for torture/execute 
> testsuite.
> Please review with the intention of committing this chnage for avr 
> target.
>
>
> 2000-03-02  Andrew Hutchinson <hutchinsonandy@aim.com>
>
>    PR target/32871
>    * config/avr/avr.c (avr_args): Add new function to set argument
>    registers in bitmap.
>    (avr_regs_to_save): Don't mark argument registers for saving.
>    (sequent_regs_live): Add bitmap argument. Use bitmap to determine
>    sequence of registers.
>    (expand_prologue): Pass bitmap of registers for sequence.
>    (expand_epilogue): Pass bitmap of registers for sequence.