[deliverable/binutils-gdb.git] / ld / emultempl / spu_ovl.S

/* Overlay manager for SPU.

   Copyright 2006 Free Software Foundation, Inc.

   This file is part of GLD, the Gnu Linker.

   GLD is free software; you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
   the Free Software Foundation; either version 2, or (at your option)
   any later version.

   GLD is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with GLD; see the file COPYING.  If not, write to the Free
   Software Foundation, 51 Franklin Street - Fifth Floor, Boston, MA
   02110-1301, USA.  */

/**
 * MFC DMA defn's.
 */
#define MFC_GET_CMD		0x40
#define MFC_MAX_DMA_SIZE	0x4000
#define MFC_TAG_UPDATE_ALL	2
#define MFC_TAG_ID		0


/**
 * Temporary register allocations.
 * These are saved/restored here.
 */
#define tab		$75
#define cgbits		$75
#define add64		$75
#define ealo		$75
#define newmask		$75
#define tagstat		$75
#define bchn		$75
#define rv1		$75

#define off		$76
#define off64		$76
#define maxsize		$76
#define oldmask		$76
#define sz		$76
#define lnkr		$76
#define rv2		$76

#define cur		$77
#define cmp		$77
#define buf		$77
#define genwi		$77
#define tagid		$77
#define cmd		$77
#define rv3		$77

#define cgshuf		$78

#define vma		$6

#define map		$7
#define osize		$7
#define cmp2		$7

#define ea64		$8
#define retval		$8

#ifdef OVLY_IRQ_SAVE
#define irqtmp		$8
#define irq_stat	$9
#endif

	.extern		_ovly_table
	.extern		_ovly_buf_table

	.text
	.align 		4
__rv_pattern:
	.word		0x00010203, 0x1c1d1e1f, 0x00010203, 0x10111213
__cg_pattern:
	.word		0x04050607, 0x80808080, 0x80808080, 0x80808080

/**
 * __ovly_return - stub for returning from overlay functions.
 *
 * inputs:
 *	$lr	link register
 *
 * outputs:
 *	$78	old partition number, to be reloaded
 *	$79	return address in old partion number
 */
	.global		__ovly_return
	.type		__ovly_return, @function

	.word		0
__ovly_return:
	shlqbyi		$78, $lr, 4
	shlqbyi		$79, $lr, 8
	biz		$78, $79
	.size		__ovly_return, . - __ovly_return

/**
 * __ovly_load - copy an overlay partion to local store.
 *
 * inputs:
 *	$78	partition number to be loaded.
 *	$79	branch target in new partition.
 *	$lr	link register, containing return addr.
 *
 * outputs:
 *	$lr	new link register, returning through __ovly_return.
 *
 * Copy a new overlay partition into local store, or return 
 * immediately if the partition is already resident.
 */
	.global		__ovly_load
	.type		__ovly_load, @function

__ovly_load:
/* Save temporary registers to stack. */
	stqd		$6, -16($sp)
	stqd		$7, -32($sp)
	stqd		$8, -48($sp)

#ifdef OVLY_IRQ_SAVE
/* Save irq state, then disable interrupts. */ 
	stqd		$9, -64($sp)
	ila		irqtmp, __ovly_irq_save
	rdch		irq_stat, $SPU_RdMachStat
	bid		irqtmp
__ovly_irq_save:
#endif

/* Set branch hint to overlay target. */
	hbr		__ovly_load_ret, $79

/* Get caller's overlay index by back chaining through stack frames.
 * Loop until end of stack (back chain all-zeros) or
 * encountered a link register we set here. */
	lqd		bchn, 0($sp)
	ila		retval, __ovly_return

__ovly_backchain_loop:
	lqd		lnkr, 16(bchn)
	lqd		bchn, 0(bchn)
	ceq		cmp, lnkr, retval
	ceqi		cmp2, bchn, 0
	or		cmp, cmp, cmp2	
	brz		cmp, __ovly_backchain_loop

/* If we reached the zero back-chain, then lnkr is bogus.  Clear the
 * part of lnkr that we use later (slot 3). */
	rotqbyi		cmp2, cmp2, 4
	andc		lnkr, lnkr, cmp2

/* Set lr = {__ovly_return, prev ovl ndx, caller return adr, callee ovl ndx}. */
	lqd		rv1, (__rv_pattern-__ovly_return+4)(retval)
	shufb		rv2, retval, lnkr, rv1
	shufb		rv3, $lr, $78, rv1
	fsmbi		rv1, 0xff
	selb		$lr, rv2, rv3, rv1

/* Branch to $79 if non-overlay */
	brz		$78, __ovly_load_restore

/* Load values from _ovly_table[$78].
 *	extern struct {
 *		u32 vma;
 *		u32 size;
 *		u32 file_offset;
 *		u32 buf;
 *	} _ovly_table[];
 */
	shli		off, $78, 4
	ila		tab, _ovly_table - 16
	lqx		vma, tab, off
	rotqbyi		buf, vma, 12

/* Load values from _ovly_buf_table[buf].
 *	extern struct {
 *		u32 mapped;
 *	} _ovly_buf_table[];
 */
	ila		tab, _ovly_buf_table
	ai		off, buf, -1
	shli		off, off, 2
	lqx		map, tab, off
	rotqby		cur, map, off

/* Branch to $79 now if overlay is already mapped.  */
	ceq		cmp, $78, cur
	brnz		cmp, __ovly_load_restore

/* Marker for profiling code.  If we get here, we are about to load
 * a new overlay.
 */
	.global		__ovly_load_event
	.type		__ovly_load_event, @function
__ovly_load_event:

/* Set _ovly_buf_table[buf].mapped = $78. */
	cwx		genwi, tab, off
	shufb		map, $78, map, genwi
	stqx		map, tab, off

/* A new partition needs to be loaded. Prepare for DMA loop. 
 * _EAR_ is the 64b base EA, filled in at run time by the
 * loader, and indicating the value for SPU executable image start.
 */
	lqd		cgshuf, (__cg_pattern-__ovly_return+4)(retval)
	rotqbyi		osize, vma, 4
	rotqbyi		sz, vma, 8
	lqa		ea64, _EAR_

__ovly_xfer_loop:
/* 64b add to compute next ea64. */
	rotqmbyi	off64, sz, -4
	cg		cgbits, ea64, off64
	shufb		add64, cgbits, cgbits, cgshuf
	addx		add64, ea64, off64
	ori		ea64, add64, 0

/* Setup DMA parameters, then issue DMA request. */
	rotqbyi		ealo, add64, 4
	ila		maxsize, MFC_MAX_DMA_SIZE
	cgt		cmp, osize, maxsize
	selb		sz, osize, maxsize, cmp
	ila		tagid, MFC_TAG_ID
	wrch		$MFC_LSA, vma
	wrch		$MFC_EAH, ea64
	wrch		$MFC_EAL, ealo
	wrch		$MFC_Size, sz
	wrch		$MFC_TagId, tagid
	ila		cmd, MFC_GET_CMD
	wrch		$MFC_Cmd, cmd

/* Increment vma, decrement size, branch back as needed. */
	a		vma, vma, sz
	sf		osize, sz, osize
	brnz		osize, __ovly_xfer_loop

/* Save app's tagmask, wait for DMA complete, restore mask. */
	rdch		oldmask, $MFC_RdTagMask
#if MFC_TAG_ID < 16
	ilh		newmask, 1 << MFC_TAG_ID
#else
	ilhu		newmask, 1 << (MFC_TAG_ID - 16)
#endif
	wrch		$MFC_WrTagMask, newmask
	ila		tagstat, MFC_TAG_UPDATE_ALL
	wrch		$MFC_WrTagUpdate, tagstat
	rdch		tagstat, $MFC_RdTagStat
	sync
	wrch		$MFC_WrTagMask, oldmask

	.global		_ovly_debug_event
	.type		_ovly_debug_event, @function
_ovly_debug_event:
/* GDB inserts debugger trap here.  */
	nop

__ovly_load_restore:
#ifdef OVLY_IRQ_SAVE
/* Conditionally re-enable interrupts. */
	andi		irq_stat, irq_stat, 1
	ila		irqtmp, __ovly_irq_restore
	binze		irq_stat, irqtmp
__ovly_irq_restore:
	lqd		$9, -64($sp)
#endif

/* Restore saved registers. */
	lqd		$8, -48($sp)
	lqd		$7, -32($sp)
	lqd		$6, -16($sp)

__ovly_load_ret:
/* Branch to target address. */
	bi		$79

	.size		__ovly_load, . - __ovly_load
Commit	Line	Data
e9f53129 AM	1	/* Overlay manager for SPU.
	2
	3	Copyright 2006 Free Software Foundation, Inc.
	4
	5	This file is part of GLD, the Gnu Linker.
	6
	7	GLD is free software; you can redistribute it and/or modify
	8	it under the terms of the GNU General Public License as published by
	9	the Free Software Foundation; either version 2, or (at your option)
	10	any later version.
	11
	12	GLD is distributed in the hope that it will be useful,
	13	but WITHOUT ANY WARRANTY; without even the implied warranty of
	14	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	15	GNU General Public License for more details.
	16
	17	You should have received a copy of the GNU General Public License
	18	along with GLD; see the file COPYING. If not, write to the Free
	19	Software Foundation, 51 Franklin Street - Fifth Floor, Boston, MA
	20	02110-1301, USA. */
	21
	22	/**
	23	* MFC DMA defn's.
	24	*/
	25	#define MFC_GET_CMD 0x40
	26	#define MFC_MAX_DMA_SIZE 0x4000
	27	#define MFC_TAG_UPDATE_ALL 2
	28	#define MFC_TAG_ID 0
	29
	30
	31	/**
	32	* Temporary register allocations.
	33	* These are saved/restored here.
	34	*/
	35	#define tab $75
	36	#define cgbits $75
	37	#define add64 $75
	38	#define ealo $75
	39	#define newmask $75
	40	#define tagstat $75
	41	#define bchn $75
	42	#define rv1 $75
	43
	44	#define off $76
	45	#define off64 $76
	46	#define maxsize $76
	47	#define oldmask $76
	48	#define sz $76
	49	#define lnkr $76
	50	#define rv2 $76
	51
	52	#define cur $77
	53	#define cmp $77
	54	#define buf $77
	55	#define genwi $77
	56	#define tagid $77
	57	#define cmd $77
	58	#define rv3 $77
	59
	60	#define cgshuf $78
	61
	62	#define vma $6
	63
	64	#define map $7
b1e37473	65	#define osize $7
e9f53129 AM	66	#define cmp2 $7
	67
	68	#define ea64 $8
	69	#define retval $8
	70
	71	#ifdef OVLY_IRQ_SAVE
	72	#define irqtmp $8
	73	#define irq_stat $9
	74	#endif
	75
	76	.extern _ovly_table
	77	.extern _ovly_buf_table
	78
	79	.text
	80	.align 4
	81	__rv_pattern:
	82	.word 0x00010203, 0x1c1d1e1f, 0x00010203, 0x10111213
	83	__cg_pattern:
	84	.word 0x04050607, 0x80808080, 0x80808080, 0x80808080
	85
	86	/**
	87	* __ovly_return - stub for returning from overlay functions.
	88	*
	89	* inputs:
	90	* $lr link register
	91	*
	92	* outputs:
	93	* $78 old partition number, to be reloaded
	94	* $79 return address in old partion number
	95	*/
	96	.global __ovly_return
	97	.type __ovly_return, @function
	98
	99	.word 0
	100	__ovly_return:
	101	shlqbyi $78, $lr, 4
	102	shlqbyi $79, $lr, 8
	103	biz $78, $79
b1e37473	104	.size __ovly_return, . - __ovly_return
e9f53129 AM	105
	106	/**
	107	* __ovly_load - copy an overlay partion to local store.
	108	*
	109	* inputs:
	110	* $78 partition number to be loaded.
	111	* $79 branch target in new partition.
	112	* $lr link register, containing return addr.
	113	*
	114	* outputs:
	115	* $lr new link register, returning through __ovly_return.
	116	*
	117	* Copy a new overlay partition into local store, or return
	118	* immediately if the partition is already resident.
	119	*/
	120	.global __ovly_load
	121	.type __ovly_load, @function
	122
	123	__ovly_load:
	124	/* Save temporary registers to stack. */
	125	stqd $6, -16($sp)
	126	stqd $7, -32($sp)
	127	stqd $8, -48($sp)
	128
	129	#ifdef OVLY_IRQ_SAVE
	130	/* Save irq state, then disable interrupts. */
	131	stqd $9, -64($sp)
	132	ila irqtmp, __ovly_irq_save
	133	rdch irq_stat, $SPU_RdMachStat
	134	bid irqtmp
	135	__ovly_irq_save:
	136	#endif
	137
	138	/* Set branch hint to overlay target. */
	139	hbr __ovly_load_ret, $79
	140
	141	/* Get caller's overlay index by back chaining through stack frames.
	142	* Loop until end of stack (back chain all-zeros) or
	143	* encountered a link register we set here. */
	144	lqd bchn, 0($sp)
	145	ila retval, __ovly_return
	146
	147	__ovly_backchain_loop:
	148	lqd lnkr, 16(bchn)
	149	lqd bchn, 0(bchn)
	150	ceq cmp, lnkr, retval
	151	ceqi cmp2, bchn, 0
	152	or cmp, cmp, cmp2
	153	brz cmp, __ovly_backchain_loop
	154
	155	/* If we reached the zero back-chain, then lnkr is bogus. Clear the
	156	* part of lnkr that we use later (slot 3). */
	157	rotqbyi cmp2, cmp2, 4
	158	andc lnkr, lnkr, cmp2
	159
	160	/* Set lr = {__ovly_return, prev ovl ndx, caller return adr, callee ovl ndx}. */
	161	lqd rv1, (__rv_pattern-__ovly_return+4)(retval)
	162	shufb rv2, retval, lnkr, rv1
	163	shufb rv3, $lr, $78, rv1
	164	fsmbi rv1, 0xff
	165	selb $lr, rv2, rv3, rv1
	166
	167	/* Branch to $79 if non-overlay */
	168	brz $78, __ovly_load_restore
169
170	/* Load values from _ovly_table[$78].
171	* extern struct {
172	* u32 vma;
173	* u32 size;
174	* u32 file_offset;
175	* u32 buf;
176	* } _ovly_table[];
177	*/
178	shli off, $78, 4
179	ila tab, _ovly_table - 16
180	lqx vma, tab, off
181	rotqbyi buf, vma, 12
182
183	/* Load values from _ovly_buf_table[buf].
184	* extern struct {
185	* u32 mapped;
186	* } _ovly_buf_table[];
187	*/
188	ila tab, _ovly_buf_table
189	ai off, buf, -1
190	shli off, off, 2
191	lqx map, tab, off
192	rotqby cur, map, off
193
194	/* Branch to $79 now if overlay is already mapped. */
195	ceq cmp, $78, cur
196	brnz cmp, __ovly_load_restore
197
b1e37473 AM	198	/* Marker for profiling code. If we get here, we are about to load
	199	* a new overlay.
	200	*/
	201	.global __ovly_load_event
	202	.type __ovly_load_event, @function
	203	__ovly_load_event:
	204
e9f53129 AM	205	/* Set _ovly_buf_table[buf].mapped = $78. */
	206	cwx genwi, tab, off
	207	shufb map, $78, map, genwi
	208	stqx map, tab, off
	209
	210	/* A new partition needs to be loaded. Prepare for DMA loop.
	211	* _EAR_ is the 64b base EA, filled in at run time by the
	212	* loader, and indicating the value for SPU executable image start.
	213	*/
	214	lqd cgshuf, (__cg_pattern-__ovly_return+4)(retval)
b1e37473	215	rotqbyi osize, vma, 4
e9f53129 AM	216	rotqbyi sz, vma, 8
	217	lqa ea64, _EAR_
	218
	219	__ovly_xfer_loop:
	220	/* 64b add to compute next ea64. */
	221	rotqmbyi off64, sz, -4
	222	cg cgbits, ea64, off64
	223	shufb add64, cgbits, cgbits, cgshuf
	224	addx add64, ea64, off64
	225	ori ea64, add64, 0
	226
	227	/* Setup DMA parameters, then issue DMA request. */
	228	rotqbyi ealo, add64, 4
	229	ila maxsize, MFC_MAX_DMA_SIZE
b1e37473 AM	230	cgt cmp, osize, maxsize
b1e37473 AM	231	selb sz, osize, maxsize, cmp
e9f53129 AM	232	ila tagid, MFC_TAG_ID
	233	wrch $MFC_LSA, vma
	234	wrch $MFC_EAH, ea64
	235	wrch $MFC_EAL, ealo
	236	wrch $MFC_Size, sz
	237	wrch $MFC_TagId, tagid
	238	ila cmd, MFC_GET_CMD
	239	wrch $MFC_Cmd, cmd
	240
	241	/* Increment vma, decrement size, branch back as needed. */
	242	a vma, vma, sz
b1e37473 AM	243	sf osize, sz, osize
b1e37473 AM	244	brnz osize, __ovly_xfer_loop
e9f53129 AM	245
	246	/* Save app's tagmask, wait for DMA complete, restore mask. */
	247	rdch oldmask, $MFC_RdTagMask
	248	#if MFC_TAG_ID < 16
	249	ilh newmask, 1 << MFC_TAG_ID
	250	#else
	251	ilhu newmask, 1 << (MFC_TAG_ID - 16)
	252	#endif
	253	wrch $MFC_WrTagMask, newmask
	254	ila tagstat, MFC_TAG_UPDATE_ALL
	255	wrch $MFC_WrTagUpdate, tagstat
	256	rdch tagstat, $MFC_RdTagStat
	257	sync
	258	wrch $MFC_WrTagMask, oldmask
	259
	260	.global _ovly_debug_event
	261	.type _ovly_debug_event, @function
	262	_ovly_debug_event:
	263	/* GDB inserts debugger trap here. */
	264	nop
	265
	266	__ovly_load_restore:
	267	#ifdef OVLY_IRQ_SAVE
	268	/* Conditionally re-enable interrupts. */
	269	andi irq_stat, irq_stat, 1
	270	ila irqtmp, __ovly_irq_restore
	271	binze irq_stat, irqtmp
	272	__ovly_irq_restore:
	273	lqd $9, -64($sp)
	274	#endif
	275
	276	/* Restore saved registers. */
	277	lqd $8, -48($sp)
	278	lqd $7, -32($sp)
	279	lqd $6, -16($sp)
	280
	281	__ovly_load_ret:
	282	/* Branch to target address. */
	283	bi $79
b1e37473 AM	284
b1e37473 AM	285	.size __ovly_load, . - __ovly_load