public inbox for gcc@gcc.gnu.org
 help / color / mirror / Atom feed
From: ibanez@sunplus.com
To: gcc@gcc.gnu.org
Subject: Backend: combine compare & arithmetic instructions
Date: Mon, 18 Jul 2005 15:19:00 -0000	[thread overview]
Message-ID: <OF75151B3A.AA7EE445-ON48257042.00541908-48257042.00541917@sunplus.com> (raw)





Thank you for so much help.
Now I can successfully combine a arith instrucion & compare.
But another problem occurs, that is our RISC machine has
logical instructions which only update the Zero flag.
Thus only eq & neq branch can be combined with.
For example

case 1 (can combine)

      or_c  dst, sr1, sr2              <-- update Z
      beq                          <-- use Z only

case 2 (can not combine)
      or_c          dst, sr1, sr2
      compare   dst, 0           <-- another compare is needed
      ble                          <-- use Z & N & V

I use a CC_Zmode to describe this.
My pattern is:

[(set (reg:CC_Z CC_REGNUM)
            (compare:CC_Z
                   (ior:SI [(match_operand:SI 1 "register_operand" "g")
                                (match_operand:SI 2 "register_operand"
"g")])
                   (const_int 0)))
        (set (match_operand:SI 0 "register_operand" "=g")
              (ior:SI [(match_dup 1) (match_dup 2)]))])]

Use gdb to trace the cc1 I saw

step1) cmpsi is invoked

 (define_expand "cmpsi"
  [(match_operand:SI 0 "register_operand" "")
   (match_operand:SI 1 "arith_operand"    "")]
  ""
  {
    compare_op0 = operands[0];
    compare_op1 = operands[1];
  }

step2)

  (define_expand "beq"
  [(set (pc)
      (if_then_else (eq (match_dup 1) (const_int 0))
                  (label_ref (match_operand 0 "" ""))
                  (pc)))]
  ""
  {
      enum machine_mode mode = SELECT_CC_MODE(EQ, cmp_op0, cmp_op1);
            /*
                  Because cmp_op0 is pseudo register, mode is always
                  the CCmode
            */
            rtx cc_reg = gen_rtx_REG(mode, CC_REGNUM);
            emit_insn(gen_rtx_SET(VOIDmode, cc_reg,
                              gen_rtx_COMPARE(mode, cmp_op0, cmp_op1)));
            /*
                  generate (set (reg:CC 33 ) (compare:CC (ior...)))

            */
            operands[1] = cc_reg;
  }
)

step3) select_cc_mode()

I set a break point in select_cc_mode() and found
it is invoked several times.

      score_select_cc_mode(enum rtx_code op, rtx x, rtx y){
            if( GET_MODE(x) == SImode && GET_CODE(x) == IOR){
                  if(op == NE || op == EQ)
                        return CC_Zmode;
                  else
                        return CCmode;
            }
            ...
      }

I also saw a
(set (reg:CC 33 ) (compare:CC (ior...)))
is changed to
(set (reg:CC_Z 33 ) (compare:CC_Z (ior...)))

Unfortunately, the combination still failed.


The rtl pattern(*.29.ce3) is


(insn:HI 10 9 11 0 main.c:71
            (set (reg:CC_Z 33 cr1)
                (compare:CC_Z
                              (ior:SI (reg:SI 4 r4 [ a ]) (const_int 16
[0x10]))
                    (const_int 0 [0x0])))
            (set (reg:SI 4 r4 [168])
                (ior:SI (reg:SI 4 r4 [ a ]) (const_int 16 [0x10]))))
        ]) 8 {bitset} (nil)
    (expr_list:REG_UNUSED (reg:CC_Z 33 cr1)
        (nil)))

(insn:HI 11 10 12 0 main.c:71 (set (reg:CC 33 cr1)
        (compare:CC (reg:SI 4 r4 [168])
            (const_int 0 [0x0]))) 42 {cmpsi_internal} (insn_list 10 (nil))
    (expr_list:REG_DEAD (reg:SI 4 r4 [168])
        (nil)))

I also trace the implementation of arm.
But it also generate a compare instruction even there is
a possibility to combine it with previous arith instructions.

I've spent so much time on it.
Can any one help me ?

                 reply	other threads:[~2005-07-18 15:19 UTC|newest]

Thread overview: [no followups] expand[flat|nested]  mbox.gz  Atom feed

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=OF75151B3A.AA7EE445-ON48257042.00541908-48257042.00541917@sunplus.com \
    --to=ibanez@sunplus.com \
    --cc=gcc@gcc.gnu.org \
    /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).