[apple/xnu.git] / osfmk / ppc / commpage / bzero_128.s

/*
 * Copyright (c) 2003 Apple Computer, Inc. All rights reserved.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
 * 
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * unlawful or unlicensed copies of an Apple operating system, or to
 * circumvent, violate, or enable the circumvention or violation of, any
 * terms of an Apple operating system software license agreement.
 * 
 * Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this file.
 * 
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 * 
 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 */

#define	ASSEMBLER
#include <sys/appleapiopts.h>
#include <ppc/asm.h>
#include <machine/cpu_capabilities.h>
#include <machine/commpage.h>

        .text
        .align	2
/*
 * WARNING: this code is written for 32-bit mode, and ported by the kernel if necessary
 * to 64-bit mode for use in the 64-bit commpage.  This "port" consists of the following
 * simple transformations:
 *      - all word compares are changed to doubleword
 *      - all "srwi[.]" opcodes are changed to "srdi[.]"                      
 * Nothing else is done.  For this to work, the following rules must be
 * carefully followed:
 *      - do not use carry or overflow
 *      - only use record mode if you are sure the results are mode-invariant
 *        for example, all "andi." and almost all "rlwinm." are fine
 *      - do not use "slwi", "slw", or "srw"
 * An imaginative programmer could break the porting model in other ways, but the above
 * are the most likely problem areas.  It is perhaps surprising how well in practice
 * this simple method works.
 */        

// **********************
// * B Z E R O _ 1 2 8  *
// **********************
//
// For 64-bit processors with a 128-byte cache line.
//
// Register use:
//		r0 = zero
//		r3 = original ptr, not changed since memset returns it
//		r4 = count of bytes to set
//		r9 = working operand ptr
// WARNING: We do not touch r2 and r10-r12, which some callers depend on.

        .align	5
bzero_128:						// void	bzero(void *b, size_t len);
        cmplwi	cr7,r4,128		// too short for DCBZ128?
        li		r0,0			// get a 0
        neg		r5,r3			// start to compute #bytes to align
        mr		r9,r3			// make copy of operand ptr (can't change r3)
        blt		cr7,Ltail		// length < 128, too short for DCBZ

// At least 128 bytes long, so compute alignment and #cache blocks.

        andi.	r5,r5,0x7F		// r5 <-  #bytes to 128-byte align
        sub		r4,r4,r5		// adjust length
        srwi	r8,r4,7			// r8 <- 128-byte chunks
        rlwinm	r4,r4,0,0x7F	// mask length down to remaining bytes
        mtctr	r8				// set up loop count
        beq		Ldcbz			// skip if already aligned (r8!=0)
        
// 128-byte align

        mtcrf	0x01,r5			// start to move #bytes to align to cr6 and cr7
        cmpwi	cr1,r8,0		// any 128-byte cache lines to 0?
        mtcrf	0x02,r5
        
        bf		31,1f			// byte?
        stb		r0,0(r9)
        addi	r9,r9,1
1:
        bf		30,2f			// halfword?
        sth		r0,0(r9)
        addi	r9,r9,2
2:
        bf		29,3f			// word?
        stw		r0,0(r9)
        addi	r9,r9,4
3:
        bf		28,4f			// doubleword?
        std		r0,0(r9)
        addi	r9,r9,8
4:
        bf		27,5f			// quadword?
        std		r0,0(r9)
        std		r0,8(r9)
        addi	r9,r9,16
5:
        bf		26,6f			// 32-byte chunk?
        std		r0,0(r9)
        std		r0,8(r9)
        std		r0,16(r9)
        std		r0,24(r9)
        addi	r9,r9,32
6:
        bf		25,7f			// 64-byte chunk?
        std		r0,0(r9)
        std		r0,8(r9)
        std		r0,16(r9)
        std		r0,24(r9)
        std		r0,32(r9)
        std		r0,40(r9)
        std		r0,48(r9)
        std		r0,56(r9)
        addi	r9,r9,64
7:
        beq		cr1,Ltail		// no chunks to dcbz128

// Loop doing 128-byte version of DCBZ instruction.
// NB: if the memory is cache-inhibited, the kernel will clear cr7
// when it emulates the alignment exception.  Eventually, we may want
// to check for this case.

Ldcbz:
        dcbz128	0,r9			// zero another 32 bytes
        addi	r9,r9,128
        bdnz	Ldcbz

// Store trailing bytes.
//		r0 = 0
//		r4 = count
//		r9 = ptr

Ltail:
        srwi.	r5,r4,4			// r5 <- 16-byte chunks to 0
        mtcrf	0x01,r4			// remaining byte count to cr7
        mtctr	r5
        beq		2f				// skip if no 16-byte chunks
1:								// loop over 16-byte chunks
        std		r0,0(r9)
        std		r0,8(r9)
        addi	r9,r9,16
        bdnz	1b
2:
        bf		28,4f			// 8-byte chunk?
        std		r0,0(r9)
        addi	r9,r9,8
4:
        bf		29,5f			// word?
        stw		r0,0(r9)
        addi	r9,r9,4
5:
        bf		30,6f			// halfword?
        sth		r0,0(r9)
        addi	r9,r9,2
6:
        bflr	31				// byte?
        stb		r0,0(r9)
        blr

	COMMPAGE_DESCRIPTOR(bzero_128,_COMM_PAGE_BZERO,kCache128+k64Bit,0, \
				kCommPageMTCRF+kCommPageBoth+kPort32to64)
Commit	Line	Data
55e303ae A	1	/*
	2	* Copyright (c) 2003 Apple Computer, Inc. All rights reserved.
	3	*
8f6c56a5	4	* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
55e303ae	5	*
8f6c56a5 A	6	* This file contains Original Code and/or Modifications of Original Code
	7	* as defined in and that are subject to the Apple Public Source License
	8	* Version 2.0 (the 'License'). You may not use this file except in
	9	* compliance with the License. The rights granted to you under the License
	10	* may not be used to create, or enable the creation or redistribution of,
	11	* unlawful or unlicensed copies of an Apple operating system, or to
	12	* circumvent, violate, or enable the circumvention or violation of, any
	13	* terms of an Apple operating system software license agreement.
	14	*
	15	* Please obtain a copy of the License at
	16	* http://www.opensource.apple.com/apsl/ and read it before using this file.
	17	*
	18	* The Original Code and all software distributed under the License are
	19	* distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
	20	* EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
	21	* INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
	22	* FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
	23	* Please see the License for the specific language governing rights and
8ad349bb	24	* limitations under the License.
8f6c56a5 A	25	*
8f6c56a5 A	26	* @APPLE_OSREFERENCE_LICENSE_HEADER_END@
55e303ae A	27	*/
	28
	29	#define ASSEMBLER
	30	#include <sys/appleapiopts.h>
	31	#include <ppc/asm.h>
	32	#include <machine/cpu_capabilities.h>
	33	#include <machine/commpage.h>
	34
	35	.text
	36	.align 2
91447636 A	37	/*
	38	* WARNING: this code is written for 32-bit mode, and ported by the kernel if necessary
	39	* to 64-bit mode for use in the 64-bit commpage. This "port" consists of the following
	40	* simple transformations:
	41	* - all word compares are changed to doubleword
	42	* - all "srwi[.]" opcodes are changed to "srdi[.]"
	43	* Nothing else is done. For this to work, the following rules must be
	44	* carefully followed:
	45	* - do not use carry or overflow
	46	* - only use record mode if you are sure the results are mode-invariant
	47	* for example, all "andi." and almost all "rlwinm." are fine
	48	* - do not use "slwi", "slw", or "srw"
	49	* An imaginative programmer could break the porting model in other ways, but the above
	50	* are the most likely problem areas. It is perhaps surprising how well in practice
	51	* this simple method works.
	52	*/
55e303ae	53
91447636 A	54	// **********************
	55	// * B Z E R O _ 1 2 8 *
	56	// **********************
55e303ae A	57	//
	58	// For 64-bit processors with a 128-byte cache line.
	59	//
	60	// Register use:
	61	// r0 = zero
	62	// r3 = original ptr, not changed since memset returns it
	63	// r4 = count of bytes to set
	64	// r9 = working operand ptr
91447636	65	// WARNING: We do not touch r2 and r10-r12, which some callers depend on.
55e303ae A	66
	67	.align 5
	68	bzero_128: // void bzero(void *b, size_t len);
	69	cmplwi cr7,r4,128 // too short for DCBZ128?
	70	li r0,0 // get a 0
	71	neg r5,r3 // start to compute #bytes to align
	72	mr r9,r3 // make copy of operand ptr (can't change r3)
	73	blt cr7,Ltail // length < 128, too short for DCBZ
	74
	75	// At least 128 bytes long, so compute alignment and #cache blocks.
	76
	77	andi. r5,r5,0x7F // r5 <- #bytes to 128-byte align
	78	sub r4,r4,r5 // adjust length
	79	srwi r8,r4,7 // r8 <- 128-byte chunks
	80	rlwinm r4,r4,0,0x7F // mask length down to remaining bytes
	81	mtctr r8 // set up loop count
	82	beq Ldcbz // skip if already aligned (r8!=0)
	83
	84	// 128-byte align
	85
	86	mtcrf 0x01,r5 // start to move #bytes to align to cr6 and cr7
	87	cmpwi cr1,r8,0 // any 128-byte cache lines to 0?
	88	mtcrf 0x02,r5
	89
	90	bf 31,1f // byte?
	91	stb r0,0(r9)
	92	addi r9,r9,1
	93	1:
	94	bf 30,2f // halfword?
	95	sth r0,0(r9)
	96	addi r9,r9,2
	97	2:
	98	bf 29,3f // word?
	99	stw r0,0(r9)
	100	addi r9,r9,4
	101	3:
	102	bf 28,4f // doubleword?
	103	std r0,0(r9)
	104	addi r9,r9,8
	105	4:
	106	bf 27,5f // quadword?
	107	std r0,0(r9)
	108	std r0,8(r9)
	109	addi r9,r9,16
	110	5:
	111	bf 26,6f // 32-byte chunk?
	112	std r0,0(r9)
	113	std r0,8(r9)
	114	std r0,16(r9)
	115	std r0,24(r9)
	116	addi r9,r9,32
	117	6:
	118	bf 25,7f // 64-byte chunk?
	119	std r0,0(r9)
	120	std r0,8(r9)
	121	std r0,16(r9)
	122	std r0,24(r9)
	123	std r0,32(r9)
	124	std r0,40(r9)
	125	std r0,48(r9)
	126	std r0,56(r9)
	127	addi r9,r9,64
	128	7:
	129	beq cr1,Ltail // no chunks to dcbz128
130
131	// Loop doing 128-byte version of DCBZ instruction.
132	// NB: if the memory is cache-inhibited, the kernel will clear cr7
133	// when it emulates the alignment exception. Eventually, we may want
134	// to check for this case.
135
136	Ldcbz:
137	dcbz128 0,r9 // zero another 32 bytes
138	addi r9,r9,128
139	bdnz Ldcbz
140
141	// Store trailing bytes.
142	// r0 = 0
143	// r4 = count
144	// r9 = ptr
145
146	Ltail:
147	srwi. r5,r4,4 // r5 <- 16-byte chunks to 0
148	mtcrf 0x01,r4 // remaining byte count to cr7
149	mtctr r5
150	beq 2f // skip if no 16-byte chunks
151	1: // loop over 16-byte chunks
152	std r0,0(r9)
153	std r0,8(r9)
154	addi r9,r9,16
155	bdnz 1b
156	2:
157	bf 28,4f // 8-byte chunk?
158	std r0,0(r9)
159	addi r9,r9,8
160	4:
161	bf 29,5f // word?
162	stw r0,0(r9)
163	addi r9,r9,4
164	5:
165	bf 30,6f // halfword?
166	sth r0,0(r9)
167	addi r9,r9,2
168	6:
169	bflr 31 // byte?
170	stb r0,0(r9)
171	blr
172
91447636 A	173	COMMPAGE_DESCRIPTOR(bzero_128,_COMM_PAGE_BZERO,kCache128+k64Bit,0, \
91447636 A	174	kCommPageMTCRF+kCommPageBoth+kPort32to64)