arch/c6x/lib/divremu.S

;; SPDX-License-Identifier: GPL-2.0-or-later
;;  Copyright 2011  Free Software Foundation, Inc.
;;  Contributed by Bernd Schmidt <bernds@codesourcery.com>.
;;

#include <linux/linkage.h>

	.text
ENTRY(__c6xabi_divremu)
	;; We use a series of up to 31 subc instructions.  First, we find
	;; out how many leading zero bits there are in the divisor.  This
	;; gives us both a shift count for aligning (shifting) the divisor
	;; to the, and the number of times we have to execute subc.

	;; At the end, we have both the remainder and most of the quotient
	;; in A4.  The top bit of the quotient is computed first and is
	;; placed in A2.

	;; Return immediately if the dividend is zero.	Setting B4 to 1
	;; is a trick to allow us to leave the following insns in the jump
	;; delay slot without affecting the result.
	mv	.s2x	A4, B1

  [b1]	lmbd	.l2	1, B4, B1
||[!b1] b	.s2	B3	; RETURN A
||[!b1] mvk	.d2	1, B4

||[!b1] zero	.s1	A5
	mv	.l1x	B1, A6
||	shl	.s2	B4, B1, B4

	;; The loop performs a maximum of 28 steps, so we do the
	;; first 3 here.
	cmpltu	.l1x	A4, B4, A2
  [!A2]	sub	.l1x	A4, B4, A4
||	shru	.s2	B4, 1, B4
||	xor	.s1	1, A2, A2

	shl	.s1	A2, 31, A2
|| [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1

	;; RETURN A may happen here (note: must happen before the next branch)
__divremu0:
	cmpgt	.l2	B1, 7, B0
|| [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
|| [b0] b	.s1	__divremu0
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
   [b1]	subc	.l1x	A4,B4,A4
|| [b1]	add	.s2	-1, B1, B1
	;; loop backwards branch happens here

	ret	.s2	B3
||	mvk	.s1	32, A1
	sub	.l1	A1, A6, A6
||	extu	.s1	A4, A6, A5
	shl	.s1	A4, A6, A4
	shru	.s1	A4, 1, A4
||	sub	.l1	A6, 1, A6
	or	.l1	A2, A4, A4
	shru	.s1	A4, A6, A4
	nop
ENDPROC(__c6xabi_divremu)