!
! Fast SH memcpy
!
! by Toshiyasu Morita (tm@netcom.com)
! hacked by J"orn Rernnecke (joern.rennecke@superh.com) ("o for o-umlaut)
! SH5 code Copyright 2002 SuperH Ltd.
!
! Entry: ARG0: destination pointer
!        ARG1: source pointer
!        ARG2: byte count
!
! Exit:  RESULT: destination pointer
!        any other registers in the range r0-r7: trashed
!
! Notes: Usually one wants to do small reads and write a longword, but
!        unfortunately it is difficult in some cases to concatanate bytes
!        into a longword on the SH, so this does a longword read and small
!        writes.
!
! This implementation makes two assumptions about how it is called:
!
! 1.: If the byte count is nonzero, the address of the last byte to be
!     copied is unsigned greater than the address of the first byte to
!     be copied.  This could be easily swapped for a signed comparison,
!     but the algorithm used needs some comparison.
!
! 2.: When there are two or three bytes in the last word of an 11-or-more
!     bytes memory chunk to b copied, the rest of the word can be read
!     without side effects.
!     This could be easily changed by increasing the minimum size of
!     a fast memcpy and the amount subtracted from r7 before L_2l_loop be 2,
!     however, this would cost a few extra cyles on average.
!     For SHmedia, the assumption is that any quadword can be read in its
!     enirety if at least one byte is included in the copy.

/* Imported into Linux kernel by Richard Curnow.  This is used to implement the
   __copy_user function in the general case, so it has to be a distinct
   function from intra-kernel memcpy to allow for exception fix-ups in the
   event that the user pointer is bad somewhere in the copy (e.g. due to
   running off the end of the vma).

   Note, this algorithm will be slightly wasteful in the case where the source
   and destination pointers are equally aligned, because the stlo/sthi pairs
   could then be merged back into single stores.  If there are a lot of cache
   misses, this is probably offset by the stall lengths on the preloads.

*/

/* NOTE : Prefetches removed and allocos guarded by synco to avoid TAKum03020
 * erratum.  The first two prefetches are nop-ed out to avoid upsetting the
 * instruction counts used in the jump address calculation.
 * */

	.section .text..SHmedia32,"ax"
	.little
	.balign 32
	.global copy_user_memcpy
	.global copy_user_memcpy_end
copy_user_memcpy:

#define LDUAQ(P,O,D0,D1) ldlo.q P,O,D0; ldhi.q P,O+7,D1
#define STUAQ(P,O,D0,D1) stlo.q P,O,D0; sthi.q P,O+7,D1
#define LDUAL(P,O,D0,D1) ldlo.l P,O,D0; ldhi.l P,O+3,D1
#define STUAL(P,O,D0,D1) stlo.l P,O,D0; sthi.l P,O+3,D1

	nop ! ld.b r3,0,r63 ! TAKum03020
	pta/l Large,tr0
	movi 25,r0
	bgeu/u r4,r0,tr0
	nsb r4,r0
	shlli r0,5,r0
	movi (L1-L0+63*32 + 1) & 0xffff,r1
	sub r1, r0, r0
L0:	ptrel r0,tr0
	add r2,r4,r5
	ptabs r18,tr1
	add r3,r4,r6
	blink tr0,r63

/* Rearranged to make cut2 safe */
	.balign 8
L4_7:	/* 4..7 byte memcpy cntd. */
	stlo.l r2, 0, r0
	or r6, r7, r6
	sthi.l r5, -1, r6
	stlo.l r5, -4, r6
	blink tr1,r63

	.balign 8
L1:	/* 0 byte memcpy */
	nop
	blink tr1,r63
	nop
	nop
	nop
	nop

L2_3:	/* 2 or 3 byte memcpy cntd. */
	st.b r5,-1,r6
	blink tr1,r63

	/* 1 byte memcpy */
	ld.b r3,0,r0
	st.b r2,0,r0
	blink tr1,r63

L8_15:	/* 8..15 byte memcpy cntd. */
	stlo.q r2, 0, r0
	or r6, r7, r6
	sthi.q r5, -1, r6
	stlo.q r5, -8, r6
	blink tr1,r63

	/* 2 or 3 byte memcpy */
	ld.b r3,0,r0
	nop ! ld.b r2,0,r63 ! TAKum03020
	ld.b r3,1,r1
	st.b r2,0,r0
	pta/l L2_3,tr0
	ld.b r6,-1,r6
	st.b r2,1,r1
	blink tr0, r63

	/* 4 .. 7 byte memcpy */
	LDUAL (r3, 0, r0, r1)
	pta L4_7, tr0
	ldlo.l r6, -4, r7
	or r0, r1, r0
	sthi.l r2, 3, r0
	ldhi.l r6, -1, r6
	blink tr0, r63

	/* 8 .. 15 byte memcpy */
	LDUAQ (r3, 0, r0, r1)
	pta L8_15, tr0
	ldlo.q r6, -8, r7
	or r0, r1, r0
	sthi.q r2, 7, r0
	ldhi.q r6, -1, r6
	blink tr0, r63

	/* 16 .. 24 byte memcpy */
	LDUAQ (r3, 0, r0, r1)
	LDUAQ (r3, 8, r8, r9)
	or r0, r1, r0
	sthi.q r2, 7, r0
	or r8, r9, r8
	sthi.q r2, 15, r8
	ldlo.q r6, -8, r7
	ldhi.q r6, -1, r6
	stlo.q r2, 8, r8
	stlo.q r2, 0, r0
	or r6, r7, r6
	sthi.q r5, -1, r6
	stlo.q r5, -8, r6
	blink tr1,r63

Large:
	! ld.b r2, 0, r63 ! TAKum03020
	pta/l  Loop_ua, tr1
	ori r3, -8, r7
	sub r2, r7, r22
	sub r3, r2, r6
	add r2, r4, r5
	ldlo.q r3, 0, r0
	addi r5, -16, r5
	movi 64+8, r27 ! could subtract r7 from that.
	stlo.q r2, 0, r0
	sthi.q r2, 7, r0
	ldx.q r22, r6, r0
	bgtu/l r27, r4, tr1

	addi r5, -48, r27
	pta/l Loop_line, tr0
	addi r6, 64, r36
	addi r6, -24, r19
	addi r6, -16, r20
	addi r6, -8, r21

Loop_line:
	! ldx.q r22, r36, r63 ! TAKum03020
	alloco r22, 32
	synco
	addi r22, 32, r22
	ldx.q r22, r19, r23
	sthi.q r22, -25, r0
	ldx.q r22, r20, r24
	ldx.q r22, r21, r25
	stlo.q r22, -32, r0
	ldx.q r22, r6,  r0
	sthi.q r22, -17, r23
	sthi.q r22,  -9, r24
	sthi.q r22,  -1, r25
	stlo.q r22, -24, r23
	stlo.q r22, -16, r24
	stlo.q r22,  -8, r25
	bgeu r27, r22, tr0

Loop_ua:
	addi r22, 8, r22
	sthi.q r22, -1, r0
	stlo.q r22, -8, r0
	ldx.q r22, r6, r0
	bgtu/l r5, r22, tr1

	add r3, r4, r7
	ldlo.q r7, -8, r1
	sthi.q r22, 7, r0
	ldhi.q r7, -1, r7
	ptabs r18,tr1
	stlo.q r22, 0, r0
	or r1, r7, r1
	sthi.q r5, 15, r1
	stlo.q r5, 8, r1
	blink tr1, r63
copy_user_memcpy_end:
	nop