From: Richard Sandiford <richard.sandiford@arm.com>
To: Robin Dapp <rdapp@linux.ibm.com>
Cc: gcc-patches@gcc.gnu.org
Subject: Re: [PATCH 4/7] ifcvt/optabs: Allow using a CC comparison for emit_conditional_move.
Date: Thu, 15 Jul 2021 21:54:12 +0100 [thread overview]
Message-ID: <mptv95b5mdn.fsf@arm.com> (raw)
In-Reply-To: <20210625160905.23786-5-rdapp@linux.ibm.com> (Robin Dapp's message of "Fri, 25 Jun 2021 18:09:02 +0200")
Robin Dapp <rdapp@linux.ibm.com> writes:
> Currently we only ever call emit_conditional_move with the comparison
> (as well as its comparands) we got from the jump. Thus, backends are
> going to emit a CC comparison for every conditional move that is being
> generated instead of re-using the existing CC.
> This, combined with emitting temporaries for each conditional move,
> causes sky-high costs for conditional moves.
>
> This patch allows to re-use a CC so the costing situation is improved a
> bit.
> ---
> gcc/ifcvt.c | 16 +++--
> gcc/optabs.c | 163 ++++++++++++++++++++++++++++++++++-----------------
> gcc/optabs.h | 1 +
> 3 files changed, 121 insertions(+), 59 deletions(-)
>
> diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c
> index ac0c142c9fe..c5b8641e2aa 100644
> --- a/gcc/ifcvt.c
> +++ b/gcc/ifcvt.c
> @@ -771,7 +771,7 @@ static int noce_try_addcc (struct noce_if_info *);
> static int noce_try_store_flag_constants (struct noce_if_info *);
> static int noce_try_store_flag_mask (struct noce_if_info *);
> static rtx noce_emit_cmove (struct noce_if_info *, rtx, enum rtx_code, rtx,
> - rtx, rtx, rtx);
> + rtx, rtx, rtx, rtx = NULL, rtx = NULL);
> static int noce_try_cmove (struct noce_if_info *);
> static int noce_try_cmove_arith (struct noce_if_info *);
> static rtx noce_get_alt_condition (struct noce_if_info *, rtx, rtx_insn **);
> @@ -1710,7 +1710,8 @@ noce_try_store_flag_mask (struct noce_if_info *if_info)
>
> static rtx
> noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code,
> - rtx cmp_a, rtx cmp_b, rtx vfalse, rtx vtrue)
> + rtx cmp_a, rtx cmp_b, rtx vfalse, rtx vtrue, rtx cc_cmp,
> + rtx rev_cc_cmp)
> {
> rtx target ATTRIBUTE_UNUSED;
> int unsignedp ATTRIBUTE_UNUSED;
> @@ -1756,9 +1757,14 @@ noce_emit_cmove (struct noce_if_info *if_info, rtx x, enum rtx_code code,
> unsignedp = (code == LTU || code == GEU
> || code == LEU || code == GTU);
>
> - target = emit_conditional_move (x, code, cmp_a, cmp_b, VOIDmode,
> - vtrue, vfalse, GET_MODE (x),
> - unsignedp);
> + if (cc_cmp != NULL_RTX && rev_cc_cmp != NULL_RTX)
> + target = emit_conditional_move (x, cc_cmp, rev_cc_cmp,
> + vtrue, vfalse, GET_MODE (x));
> + else
> + target = emit_conditional_move (x, code, cmp_a, cmp_b, VOIDmode,
> + vtrue, vfalse, GET_MODE (x),
> + unsignedp);
It might make sense to move:
/* Don't even try if the comparison operands are weird
except that the target supports cbranchcc4. */
if (! general_operand (cmp_a, GET_MODE (cmp_a))
|| ! general_operand (cmp_b, GET_MODE (cmp_b)))
{
if (!have_cbranchcc4
|| GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC
|| cmp_b != const0_rtx)
return NULL_RTX;
}
into the “else” arm, since it seems odd to be checking cmp_a and cmp_b
when we're not going to use them. Looks like the later call to
emit_conditional_move should get the same treatment.
> +
> if (target)
> return target;
>
> diff --git a/gcc/optabs.c b/gcc/optabs.c
> index 62a6bdb4c59..6bf486b9b50 100644
> --- a/gcc/optabs.c
> +++ b/gcc/optabs.c
> @@ -52,6 +52,8 @@ static void prepare_float_lib_cmp (rtx, rtx, enum rtx_code, rtx *,
> static rtx expand_unop_direct (machine_mode, optab, rtx, rtx, int);
> static void emit_libcall_block_1 (rtx_insn *, rtx, rtx, rtx, bool);
>
> +static rtx emit_conditional_move (rtx, rtx, rtx, rtx, machine_mode);
> +
> /* Debug facility for use in GDB. */
> void debug_optab_libfuncs (void);
> \f
> @@ -4747,7 +4749,6 @@ emit_conditional_move (rtx target, enum rtx_code code, rtx op0, rtx op1,
> machine_mode mode, int unsignedp)
> {
> rtx comparison;
> - rtx_insn *last;
> enum insn_code icode;
> enum rtx_code reversed;
>
> @@ -4774,6 +4775,7 @@ emit_conditional_move (rtx target, enum rtx_code code, rtx op0, rtx op1,
> /* get_condition will prefer to generate LT and GT even if the old
> comparison was against zero, so undo that canonicalization here since
> comparisons against zero are cheaper. */
> +
> if (code == LT && op1 == const1_rtx)
> code = LE, op1 = const0_rtx;
> else if (code == GT && op1 == constm1_rtx)
> @@ -4782,17 +4784,29 @@ emit_conditional_move (rtx target, enum rtx_code code, rtx op0, rtx op1,
> if (cmode == VOIDmode)
> cmode = GET_MODE (op0);
>
> - enum rtx_code orig_code = code;
> + /* If the first source operand is constant and the second is not, swap
> + it into the second. In that case we also need to reverse the
> + comparison. It is possible, though, that the conditional move
> + will not expand with operands in this order, so we might also need
> + to revert to the original comparison and operand order. */
Why's that the case though? The swapped form is the canonical one,
so it's the one that the target ought to accept.
Thanks,
Richard
> +
> + rtx rev_comparison = NULL_RTX;
> bool swapped = false;
> - if (swap_commutative_operands_p (op2, op3)
> - && ((reversed = reversed_comparison_code_parts (code, op0, op1, NULL))
> - != UNKNOWN))
> +
> + code = unsignedp ? unsigned_condition (code) : code;
> + comparison = simplify_gen_relational (code, VOIDmode, cmode, op0, op1);
> +
> + if ((reversed = reversed_comparison_code_parts (code, op0, op1, NULL))
> + != UNKNOWN)
> {
> - std::swap (op2, op3);
> - code = reversed;
> - swapped = true;
> + reversed = unsignedp ? unsigned_condition (reversed) : reversed;
> + rev_comparison = simplify_gen_relational (reversed, VOIDmode, cmode,
> + op0, op1);
> }
>
> + if (swap_commutative_operands_p (op2, op3) && reversed != UNKNOWN)
> + swapped = true;
> +
> if (mode == VOIDmode)
> mode = GET_MODE (op2);
>
> @@ -4804,58 +4818,99 @@ emit_conditional_move (rtx target, enum rtx_code code, rtx op0, rtx op1,
> if (!target)
> target = gen_reg_rtx (mode);
>
> - for (int pass = 0; ; pass++)
> + if (comparison && COMPARISON_P (comparison))
> + prepare_cmp_insn (XEXP (comparison, 0), XEXP (comparison, 1),
> + GET_CODE (comparison), NULL_RTX, unsignedp, OPTAB_WIDEN,
> + &comparison, &cmode);
> + else
> + return NULL_RTX;
> +
> + if (rev_comparison && COMPARISON_P (rev_comparison))
> + prepare_cmp_insn (XEXP (rev_comparison, 0), XEXP (rev_comparison, 1),
> + GET_CODE (rev_comparison), NULL_RTX,
> + unsignedp, OPTAB_WIDEN, &rev_comparison, &cmode);
> +
> + if (!swapped)
> + return emit_conditional_move (target, comparison, rev_comparison,
> + op2, op3, mode);
> + else
> + return emit_conditional_move (target, rev_comparison, comparison,
> + op3, op2, mode);
> +}
> +
> +/* Helper function for emitting a conditional move. Given a COMPARISON
> + and a reversed REV_COMPARISON it will try to expand a conditional move
> + with COMPARISON first and try with REV_COMPARISON if that fails. */
> +
> +rtx
> +emit_conditional_move (rtx target, rtx comparison, rtx rev_comparison,
> + rtx op2, rtx op3, machine_mode mode)
> +{
> +
> + rtx res = emit_conditional_move (target, comparison, op2, op3, mode);
> +
> + if (res != NULL_RTX)
> + return res;
> +
> + return emit_conditional_move (target, rev_comparison, op3, op2, mode);
> +}
> +
> +/* Helper for emitting a conditional move. */
> +
> +static rtx
> +emit_conditional_move (rtx target, rtx comparison,
> + rtx op2, rtx op3, machine_mode mode)
> +{
> + rtx_insn *last;
> + enum insn_code icode;
> +
> + if (comparison == NULL_RTX || !COMPARISON_P (comparison))
> + return NULL_RTX;
> +
> + /* If the two source operands are identical, that's just a move. */
> + if (rtx_equal_p (op2, op3))
> {
> - code = unsignedp ? unsigned_condition (code) : code;
> - comparison = simplify_gen_relational (code, VOIDmode, cmode, op0, op1);
> + if (!target)
> + target = gen_reg_rtx (mode);
>
> - /* We can get const0_rtx or const_true_rtx in some circumstances. Just
> - punt and let the caller figure out how best to deal with this
> - situation. */
> - if (COMPARISON_P (comparison))
> - {
> - saved_pending_stack_adjust save;
> - save_pending_stack_adjust (&save);
> - last = get_last_insn ();
> - do_pending_stack_adjust ();
> - machine_mode cmpmode = cmode;
> - prepare_cmp_insn (XEXP (comparison, 0), XEXP (comparison, 1),
> - GET_CODE (comparison), NULL_RTX, unsignedp,
> - OPTAB_WIDEN, &comparison, &cmpmode);
> - if (comparison)
> - {
> - class expand_operand ops[4];
> + emit_move_insn (target, op3);
> + return target;
> + }
>
> - create_output_operand (&ops[0], target, mode);
> - create_fixed_operand (&ops[1], comparison);
> - create_input_operand (&ops[2], op2, mode);
> - create_input_operand (&ops[3], op3, mode);
> - if (maybe_expand_insn (icode, 4, ops))
> - {
> - if (ops[0].value != target)
> - convert_move (target, ops[0].value, false);
> - return target;
> - }
> - }
> - delete_insns_since (last);
> - restore_pending_stack_adjust (&save);
> - }
> + if (mode == VOIDmode)
> + mode = GET_MODE (op2);
>
> - if (pass == 1)
> - return NULL_RTX;
> + icode = direct_optab_handler (movcc_optab, mode);
>
> - /* If the preferred op2/op3 order is not usable, retry with other
> - operand order, perhaps it will expand successfully. */
> - if (swapped)
> - code = orig_code;
> - else if ((reversed = reversed_comparison_code_parts (orig_code, op0, op1,
> - NULL))
> - != UNKNOWN)
> - code = reversed;
> - else
> - return NULL_RTX;
> - std::swap (op2, op3);
> + if (icode == CODE_FOR_nothing)
> + return NULL_RTX;
> +
> + if (!target)
> + target = gen_reg_rtx (mode);
> +
> + saved_pending_stack_adjust save;
> + save_pending_stack_adjust (&save);
> + last = get_last_insn ();
> + do_pending_stack_adjust ();
> +
> + class expand_operand ops[4];
> +
> + create_output_operand (&ops[0], target, mode);
> + create_fixed_operand (&ops[1], comparison);
> + create_input_operand (&ops[2], op2, mode);
> + create_input_operand (&ops[3], op3, mode);
> +
> + if (maybe_expand_insn (icode, 4, ops))
> + {
> + if (ops[0].value != target)
> + convert_move (target, ops[0].value, false);
> + return target;
> }
> +
> + delete_insns_since (last);
> + restore_pending_stack_adjust (&save);
> +
> + return NULL_RTX;
> }
>
>
> diff --git a/gcc/optabs.h b/gcc/optabs.h
> index 3bbceff92d9..f853b93f37f 100644
> --- a/gcc/optabs.h
> +++ b/gcc/optabs.h
> @@ -281,6 +281,7 @@ extern void emit_indirect_jump (rtx);
> /* Emit a conditional move operation. */
> rtx emit_conditional_move (rtx, enum rtx_code, rtx, rtx, machine_mode,
> rtx, rtx, machine_mode, int);
> +rtx emit_conditional_move (rtx, rtx, rtx, rtx, rtx, machine_mode);
>
> /* Emit a conditional negate or bitwise complement operation. */
> rtx emit_conditional_neg_or_complement (rtx, rtx_code, machine_mode, rtx,
next prev parent reply other threads:[~2021-07-15 20:54 UTC|newest]
Thread overview: 31+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-06-25 16:08 [PATCH 0/7] ifcvt: Convert multiple Robin Dapp
2021-06-25 16:08 ` [PATCH 1/7] ifcvt: Check if cmovs are needed Robin Dapp
2021-07-15 20:10 ` Richard Sandiford
2021-07-22 12:06 ` Robin Dapp
2021-07-26 19:08 ` Richard Sandiford
2021-09-15 8:39 ` Robin Dapp
2021-10-14 8:45 ` Richard Sandiford
2021-10-14 14:20 ` Robin Dapp
2021-10-14 14:32 ` Richard Sandiford
2021-10-18 11:40 ` Robin Dapp
2021-11-03 8:55 ` Robin Dapp
2021-11-05 15:33 ` Richard Sandiford
2021-11-12 13:00 ` Robin Dapp
2021-11-30 16:36 ` Richard Sandiford
2021-06-25 16:09 ` [PATCH 2/7] ifcvt: Allow constants for noce_convert_multiple Robin Dapp
2021-07-15 20:25 ` Richard Sandiford
2021-06-25 16:09 ` [PATCH 3/7] ifcvt: Improve costs handling " Robin Dapp
2021-07-15 20:42 ` Richard Sandiford
2021-07-22 12:07 ` Robin Dapp
2021-07-26 19:10 ` Richard Sandiford
2021-06-25 16:09 ` [PATCH 4/7] ifcvt/optabs: Allow using a CC comparison for emit_conditional_move Robin Dapp
2021-07-15 20:54 ` Richard Sandiford [this message]
2021-07-22 12:07 ` Robin Dapp
2021-07-26 19:31 ` Richard Sandiford
2021-07-27 20:49 ` Robin Dapp
2021-08-06 12:14 ` Richard Sandiford
2021-06-25 16:09 ` [PATCH 5/7] ifcvt: Try re-using CC for conditional moves Robin Dapp
2021-07-22 12:12 ` Robin Dapp
2021-06-25 16:09 ` [PATCH 6/7] testsuite/s390: Add tests for noce_convert_multiple Robin Dapp
2021-06-25 16:09 ` [PATCH 7/7] s390: Increase costs for load on condition and change movqicc expander Robin Dapp
2021-07-13 12:42 ` [PATCH 0/7] ifcvt: Convert multiple Robin Dapp
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=mptv95b5mdn.fsf@arm.com \
--to=richard.sandiford@arm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=rdapp@linux.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).