This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] PowerPC fsel PR5217
- From: Geoff Keating <geoffk at geoffk dot org>
- To: dje at watson dot ibm dot com
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Sun, 3 Feb 2002 21:29:02 -0800
- Subject: Re: [PATCH] PowerPC fsel PR5217
- References: <200202040514.AAA30458@makai.watson.ibm.com>
- Reply-to: Geoff Keating <geoffk at redhat dot com>
> cc: gcc-patches@gcc.gnu.org
> Date: Mon, 04 Feb 2002 00:14:23 -0500
> From: David Edelsohn <dje@watson.ibm.com>
>
> >>>>> Geoff Keating writes:
>
> Geoff> tmp = gen_rtx_fmt_ee (code, GET_MODE (if_info->cond), cmp_a, cmp_b);
> Geoff> tmp = gen_rtx_IF_THEN_ELSE (GET_MODE (x), tmp, vtrue, vfalse);
> Geoff> tmp = gen_rtx_SET (VOIDmode, x, tmp);
>
> Geoff> which generates an IF_THEN_ELSE with the same mode as the destination
> Geoff> of the SET.
>
> combine.c uses VOIDmode and cselib.c appears to use the mode of
> the source:
>
> src = gen_rtx_IF_THEN_ELSE (GET_MODE (src), cond, src, dest);
>
> which is what my pattern matches.
Um... In your patch, you had
***************
*** 5084,5090 ****
(define_insn "*fseldfsf4"
[(set (match_operand:SF 0 "gpc_reg_operand" "=f")
! (if_then_else:SF (ge (match_operand:DF 1 "gpc_reg_operand" "f")
(match_operand:DF 4 "zero_fp_constant" "F"))
(match_operand:SF 2 "gpc_reg_operand" "f")
(match_operand:SF 3 "gpc_reg_operand" "f")))]
--- 5084,5090 ----
(define_insn "*fseldfsf4"
[(set (match_operand:SF 0 "gpc_reg_operand" "=f")
! (if_then_else:DF (ge (match_operand:DF 1 "gpc_reg_operand" "f")
(match_operand:DF 4 "zero_fp_constant" "F"))
(match_operand:SF 2 "gpc_reg_operand" "f")
(match_operand:SF 3 "gpc_reg_operand" "f")))]
In this case, 'src' is '(match_operand:SF 2 "gpc_reg_operand" "f")',
and has mode SFmode.
> So the question is which way do we want to canonicalize this?
> Then we can document it and match the decision in the PowerPC port.
We certainly want the mode of the IF_THEN_ELSE to be the mode of the
value it produces; this is how almost all other RTL is done. VOIDmode
would just be annoying.
--
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>