mainline bootstrap broken for powerpc64-unknown-linux-gnu

mainline bootstrap broken for powerpc64-unknown-linux-gnu

[Janis Johnson]

Bootstrap for powerpc64-unknown-linux-gnu broken with yesterday's
changes to include darwin.md in rs6000.md.  This:

(define_insn "*call_nonlocal_darwin64"
  [(call (mem:SI (match_operand:DI 0 "symbol_ref_operand" "s,s"))
         (match_operand 1 "" "g,g"))
   (use (match_operand:SI 2 "immediate_operand" "O,n"))
   (clobber (match_scratch:SI 3 "=l,l"))]
  "(DEFAULT_ABI == ABI_DARWIN)
   && (INTVAL (operands[2]) & CALL_LONG) == 0"
{
#if TARGET_MACHO
  return output_call(insn, operands, 0, 2);
#endif
}

turns into this:

static const char *
output_3 (rtx *operands ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED)
{
{
#if TARGET_MACHO
  return output_call(insn, operands, 0, 2);
#endif
}
}

which gets a warning about falling off the end of the function.  There's
another one that's similar.

Janis

Re: mainline bootstrap broken for powerpc64-unknown-linux-gnu

[Stan Shebs]

Janis Johnson wrote:

>Bootstrap for powerpc64-unknown-linux-gnu broken with yesterday's
>changes to include darwin.md in rs6000.md.  This:
>
>(define_insn "*call_nonlocal_darwin64"
>  [(call (mem:SI (match_operand:DI 0 "symbol_ref_operand" "s,s"))
>         (match_operand 1 "" "g,g"))
>   (use (match_operand:SI 2 "immediate_operand" "O,n"))
>   (clobber (match_scratch:SI 3 "=l,l"))]
>  "(DEFAULT_ABI == ABI_DARWIN)
>   && (INTVAL (operands[2]) & CALL_LONG) == 0"
>{
>#if TARGET_MACHO
>  return output_call(insn, operands, 0, 2);
>#endif
>}
>
>turns into this:
>
>static const char *
>output_3 (rtx *operands ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED)
>{
>{
>#if TARGET_MACHO
>  return output_call(insn, operands, 0, 2);
>#endif
>}
>}
>
>which gets a warning about falling off the end of the function.  There's
>another one that's similar.
>
My fault, David already reverted.

Stan

Re: mainline bootstrap broken for powerpc64-unknown-linux-gnu

[Zack Weinberg]

Janis Johnson <janis187@us.ibm.com> writes:

> Bootstrap for powerpc64-unknown-linux-gnu broken with yesterday's
> changes to include darwin.md in rs6000.md.  This:
>
> (define_insn "*call_nonlocal_darwin64"
>   [(call (mem:SI (match_operand:DI 0 "symbol_ref_operand" "s,s"))
>          (match_operand 1 "" "g,g"))
>    (use (match_operand:SI 2 "immediate_operand" "O,n"))
>    (clobber (match_scratch:SI 3 "=l,l"))]
>   "(DEFAULT_ABI == ABI_DARWIN)
>    && (INTVAL (operands[2]) & CALL_LONG) == 0"
> {
> #if TARGET_MACHO
>   return output_call(insn, operands, 0, 2);
> #endif
> }
...

I would like to suggest that the correct fix for this is to add
TARGET_MACHO at the beginning of the C condition, and remove the
#ifdef.  This will make the pattern fail to match on ELF/XCOFF
targets, which is the right thing, and since TARGET_MACHO is a
compile-time constant, genconditions will exclude the pattern
from the generated insn-*.c files.

zw