This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: LRA and splitters
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Mike Stump <mikestump at comcast dot net>
- Cc: Vladimir Makarov <vmakarov at redhat dot com>, GCC Development <gcc at gcc dot gnu dot org>
- Date: Thu, 8 May 2014 19:24:46 +0200
- Subject: Re: LRA and splitters
- Authentication-results: sourceware.org; auth=none
- References: <85063EF0-9160-4DD4-A24F-19ACED230C83 at comcast dot net>
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
On Thu, May 08, 2014 at 10:15:06AM -0700, Mike Stump wrote:
> So, I was wondering about patterns like:
>
> (define_insn_and_split "*setcc_di_1"
> [(set (match_operand:DI 0 "register_operand" "=q")
> (match_operator:DI 1 "ix86_comparison_operator"
> [(reg FLAGS_REG) (const_int 0)]))]
> "TARGET_64BIT && !TARGET_PARTIAL_REG_STALL"
> "#"
> "&& reload_completed"
> [(set (match_dup 2) (match_dup 1))
> (set (match_dup 0) (zero_extend:DI (match_dup 2)))]
> {
> PUT_MODE (operands[1], QImode);
> operands[2] = gen_lowpart (QImode, operands[0]);
> })
>
> on the x86. X86 is an always LRA port, and with LRA it seems reload_completed is always 0, and so that disables these splitters?
No. reload_completed flag is set after reload or LRA has completed, see
lra.c. It is only reload_in_progress flag that is never true for LRA, where
it has lra_in_progress instead.
Jakub