https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116781
--- Comment #3 from Denis Chertykov <denisc at gcc dot gnu.org> ---
(In reply to Georg-Johann Lay from comment #2)
> Created attachment 59602 [details]
> pr116781-gjl.diff
>
> (In reply to Denis Chertykov from comment #1)
> > Probably we have a wring definition of "*tablejump_split"
> >
> > The patch:
> >
> > diff --git a/gcc/config/avr/avr.md b/gcc/config/avr/avr.md
> > index aae8a696a63..a26309650f2 100644
> > --- a/gcc/config/avr/avr.md
> > +++ b/gcc/config/avr/avr.md
> > @@ -7653,7 +7653,7 @@
> > (unspec:HI [(match_operand:HI 0 "register_operand" "!z,*r,z")]
> > UNSPEC_INDEX_JMP))
> > (use (label_ref (match_operand 1 "" "")))
> > - (clobber (match_dup 0))
> > + (clobber (match_scratch:HI 2 "=X,X,0"))
> > (clobber (const_int 0))]
> > "!AVR_HAVE_EIJMP_EICALL"
> > "#"
> > @@ -7759,7 +7759,7 @@
> > (parallel [(set (pc)
> > (unspec:HI [(match_dup 7)] UNSPEC_INDEX_JMP))
> > (use (label_ref (match_dup 3)))
> > - (clobber (match_dup 7))
> > + (clobber (scratch:HI))
> > (clobber (match_dup 8))])]
> > ""
> > {
>
> I am not sure whether that is expressing what is being clobbered (operand 0).
>
> The casesi expander must clobber the registers for operand 7 which it is
> cooking up, so the casesi expander looks correct to me.
>
> The tablejump insn for the 2-byte PC case would go something like:
>
> diff --git a/gcc/config/avr/avr.md b/gcc/config/avr/avr.md
> index 550b01b36fb..7ba43030570 100644
> --- a/gcc/config/avr/avr.md
> +++ b/gcc/config/avr/avr.md
> @@ -7713,7 +7713,7 @@ (define_insn_and_split "*tablejump_split"
> (unspec:HI [(match_operand:HI 0 "register_operand" "!z,*r,z")]
> UNSPEC_INDEX_JMP))
> (use (label_ref (match_operand 1 "" "")))
> - (clobber (match_dup 0))
> + (clobber (match_operand:HI 2 "register_operand" "=0,0,0"))
> (clobber (const_int 0))]
> "!AVR_HAVE_EIJMP_EICALL"
> "#"
> @@ -7722,7 +7722,7 @@ (define_insn_and_split "*tablejump_split"
> (unspec:HI [(match_dup 0)]
> UNSPEC_INDEX_JMP))
> (use (label_ref (match_dup 1)))
> - (clobber (match_dup 0))
> + (clobber (match_dup 2))
IMHO: here we can have a `(clobber (match_dup 0))' because reload_completed
(as you mentioned at the end of message).
> (clobber (const_int 0))
> (clobber (reg:CC REG_CC))])]
> ""
> @@ -7733,7 +7733,7 @@ (define_insn "*tablejump"
> (unspec:HI [(match_operand:HI 0 "register_operand" "!z,*r,z")]
> UNSPEC_INDEX_JMP))
> (use (label_ref (match_operand 1 "" "")))
> - (clobber (match_dup 0))
> + (clobber (match_operand:HI 2 "register_operand" "=0,0,0"))
> (clobber (const_int 0))
> (clobber (reg:CC REG_CC))]
> "!AVR_HAVE_EIJMP_EICALL && reload_completed"
>
> The reason is that it has to clobber regs it already has; the regs are NOT
> provided by reload.
>
> For the split pattern it should not matter whether (match_dup 0) or
> (match_dup 2) is used because they must be the same.
I'm agree with your patch.
