combine: Add added_notes_insn

This patch makes combine reconsider insns it added notes to.  This
matters for example if the note is a REG_DEAD; without the note the
setter of the register has to be kept around in the result of
combinations, so it cannot be a 2->1 combination, and the cost of
the result is higher than without that extra set, so try_combine may
refuse the combination with the set, but allow it without the set.

This fixes a regression for powerpc: pr69946.c has started to fail
after the bitfield expansion changes.  GCC used to generate

        lwz 3,0(9)
        rlwinm 3,3,12,20,23
        ori 3,3,0x11
        rotldi 3,3,52
        bl bar

but now it does

        lwz 3,0(9)
        rldicr 3,3,32,3
        srdi 3,3,48
        ori 3,3,0x110
        sldi 3,3,48
        bl bar

(an instruction too many).  After this patch it is

        lwz 3,0(9)
        rlwinm 3,3,16,16,19
        ori 3,3,0x110
        sldi 3,3,48
        bl bar

(the testcase still does not pass, it looks for very specific insns).


	* combine.c (added_notes_insn): New.
	(try_combine): Handle added_notes_insn like added_links_insn.
	Rewrite return value code.
	(distribute_notes): Set added_notes_insn to the earliest insn we added
	a note to.

From-SVN: r254875
This commit is contained in:
Segher Boessenkool 2017-11-17 15:53:29 +01:00 committed by Segher Boessenkool
parent 58b46683bd
commit f1b7bc164c
2 changed files with 33 additions and 7 deletions

View File

@ -1,3 +1,11 @@
2017-11-17 Segher Boessenkool <segher@kernel.crashing.org>
* combine.c (added_notes_insn): New.
(try_combine): Handle added_notes_insn like added_links_insn.
Rewrite return value code.
(distribute_notes): Set added_notes_insn to the earliest insn we added
a note to.
2017-11-17 Segher Boessenkool <segher@kernel.crashing.org>
PR rtl-optimization/82621

View File

@ -302,6 +302,10 @@ static HARD_REG_SET newpat_used_regs;
static rtx_insn *added_links_insn;
/* And similarly, for notes. */
static rtx_insn *added_notes_insn;
/* Basic block in which we are performing combines. */
static basic_block this_basic_block;
static bool optimize_this_for_speed_p;
@ -2790,6 +2794,7 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
std::swap (i1, i2);
added_links_insn = 0;
added_notes_insn = 0;
/* First check for one important special case that the code below will
not handle. Namely, the case where I1 is zero, I2 is a PARALLEL
@ -4752,12 +4757,13 @@ try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0,
combine_successes++;
undo_commit ();
if (added_links_insn
&& (newi2pat == 0 || DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (i2))
&& DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (i3))
return added_links_insn;
else
return newi2pat ? i2 : i3;
rtx_insn *ret = newi2pat ? i2 : i3;
if (added_links_insn && DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (ret))
ret = added_links_insn;
if (added_notes_insn && DF_INSN_LUID (added_notes_insn) < DF_INSN_LUID (ret))
ret = added_notes_insn;
return ret;
}
/* Get a marker for undoing to the current state. */
@ -14628,10 +14634,22 @@ distribute_notes (rtx notes, rtx_insn *from_insn, rtx_insn *i3, rtx_insn *i2,
{
XEXP (note, 1) = REG_NOTES (place);
REG_NOTES (place) = note;
/* Set added_notes_insn to the earliest insn we added a note to. */
if (added_notes_insn == 0
|| DF_INSN_LUID (added_notes_insn) > DF_INSN_LUID (place))
added_notes_insn = place;
}
if (place2)
add_shallow_copy_of_reg_note (place2, note);
{
add_shallow_copy_of_reg_note (place2, note);
/* Set added_notes_insn to the earliest insn we added a note to. */
if (added_notes_insn == 0
|| DF_INSN_LUID (added_notes_insn) > DF_INSN_LUID (place2))
added_notes_insn = place2;
}
}
}