Re: [Patch] 1/3 Add new builtin __builtin_flush_icache().

[Richard Sandiford <rsandifo at nildram dot co dot uk>]

David Daney <ddaney@avtrex.com> writes:
> +/* Expand a call to _builtin___clear_cache.  If the target does not
> +   support the operation, return NULL_RTX so it expands as a call. */
> +static rtx
> +expand_builtin_clear_cache (tree exp)
> +{
> +  tree begin, end;
> +  rtx begin_rtx, end_rtx;
> +
> +  if (!targetm.builtin_clear_cache_inline_p())
> +    return NULL_RTX;
> +
> +  /* We must not expand to a library call if
> +     __builtin_clear_cache_inline_p() returns true.  If we did, the
> +     fallback library function in libgcc would expand to a call to
> +     _builtin___clear_cache() and we would recurse infinitely.  */
> +  if (!validate_arglist (exp, POINTER_TYPE, POINTER_TYPE, VOID_TYPE))

This comment seems to belong above the earlier code.  However, I'm only
really replying because...

> +  /* Evaluate BEGIN and END. */
> +  begin_rtx = expand_expr (begin, NULL_RTX, Pmode, EXPAND_NORMAL);
> +  begin_rtx = convert_memory_address (Pmode, begin_rtx);
> +
> +  end_rtx = expand_expr (end, NULL_RTX, Pmode, EXPAND_NORMAL);
> +  end_rtx = convert_memory_address (Pmode, end_rtx);
> +
> +#ifdef HAVE_clear_cache
> +  if (HAVE_clear_cache)
> +    emit_insn (gen_clear_cache (begin_rtx, end_rtx));
> +#endif

When I said that the switch from general_operand to pmode_register_operand
would need target-independent changes, I meant it would need changes to
this bit of code.  I think you need to check the predicates of each
clear_cache operand and force the operand into a register if it doesn't
match.  Something like:

  if (HAVE_clear_cache)
    {
      icode = CODE_FOR_clear_cache;
      if (!insn_data[icode].operand[0].predicate (begin_rtx, Pmode))
	begin_rtx = copy_to_mode_reg (Pmode, begin_rtx);
      if (!insn_data[icode].operand[1].predicate (end_rtx, Pmode))
	end_rtx = copy_to_mode_reg (Pmode, end_rtx);
      emit_insn (gen_clear_cache (begin_rtx, end_rtx));
    }

(It's unfortunate that this construct has to be copied so often,
but that's the way it is.)

As it stands, the code only seems to guarantee that the operand
is an address operand, whereas the predicates in the MIPS patch
claim that it will be a register.

Richard