[gcc r13-2] i386: Improve ix86_expand_int_movcc

[gcc r13-2] i386: Improve ix86_expand_int_movcc

[Jakub Jelinek]

https://gcc.gnu.org/g:cb1758d95cc12e3f2d6c0e295288e4867c2e7197

commit r13-2-gcb1758d95cc12e3f2d6c0e295288e4867c2e7197
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Thu Apr 28 17:41:49 2022 +0200

    i386: Improve ix86_expand_int_movcc
    
    When working on PR105338, I've noticed that in some cases we emit
    unnecessarily long sequence which has then higher seq_cost than necessary.
    
    E.g. when ix86_expand_int_movcc is called with
    operands[0] (reg/v:SI 83 [ i ])
    operands[1] (eq (reg/v:SI 83 [ i ]) (const_int 0 [0]))
    operands[2] (reg/v:SI 83 [ i ])
    operands[3] (const_int -2 [0xfffffffffffffffe])
    i.e. r83 = r83 == 0 ? r83 : -2 which with my PR105338 patch is equivalent to
    r83 = r83 == 0 ? 0 : -2, we emit:
    (insn 24 0 25 (set (reg:CC 17 flags)
            (compare:CC (reg/v:SI 83 [ i ])
                (const_int 1 [0x1]))) 11 {*cmpsi_1}
         (nil))
    (insn 25 24 26 (parallel [
                (set (reg:SI 85)
                    (if_then_else:SI (ltu:SI (reg:CC 17 flags)
                            (const_int 0 [0]))
                        (const_int -1 [0xffffffffffffffff])
                        (const_int 0 [0])))
                (clobber (reg:CC 17 flags))
            ]) 1192 {*x86_movsicc_0_m1}
         (nil))
    (insn 26 25 27 (set (reg:SI 85)
            (not:SI (reg:SI 85))) 683 {*one_cmplsi2_1}
         (nil))
    (insn 27 26 28 (parallel [
                (set (reg:SI 85)
                    (and:SI (reg:SI 85)
                        (const_int -2 [0xfffffffffffffffe])))
                (clobber (reg:CC 17 flags))
            ]) 533 {*andsi_1}
         (nil))
    (insn 28 27 0 (set (reg/v:SI 83 [ i ])
            (reg:SI 85)) 81 {*movsi_internal}
         (nil))
    which has seq_cost (seq, true) 24.  But it could have just cost 20
    if we didn't decide to use a fresh temporary r85 and used r83 instead
    - we could avoid the copy at the end.
    The reason for it is in the 2 reg_overlap_mentioned_p calls,
    the destination (out) indeed overlaps op0 - it is the same register,
    but I don't see why that is a problem, this is in a code path where
    we've already called
    ix86_expand_carry_flag_compare (code, op0, op1, &compare_op)
    earlier, so the fact that we've out overlaps op0 or op1 shouldn't matter
    because insn 24 above is already emitted, we should just care if
    it overlaps whatever we got from that ix86_expand_carry_flag_compare
    call, i.e. compare_op, otherwise we can overwrite out just fine;
    we also know at that point that the last 2 operands of ?: are constants.
    
    2022-04-28  Jakub Jelinek  <jakub@redhat.com>
    
            * config/i386/i386-expand.cc (ix86_expand_int_movcc): Create a
            temporary only if out overlaps compare_op, not when it overlaps
            op0 or op1.

Diff:
---
 gcc/config/i386/i386-expand.cc | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index 31780b6daf7..bc806ffa283 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -3224,8 +3224,7 @@ ix86_expand_int_movcc (rtx operands[])
 		}
 	      diff = ct - cf;
 
-	      if (reg_overlap_mentioned_p (out, op0)
-		  || reg_overlap_mentioned_p (out, op1))
+	      if (reg_overlap_mentioned_p (out, compare_op))
 		tmp = gen_reg_rtx (mode);
 
 	      if (mode == DImode)