1
linux/arch/arm/lib/copy_template.S

256 lines
5.7 KiB
ArmAsm
Raw Normal View History

[ARM] 2947/1: copy template with new memcpy/memmove Patch from Nicolas Pitre This patch provides a new implementation for optimized memory copy functions on ARM. It is made of two levels: a template that consists of the core copy code and separate files that define macros to be used with the core code depending on the type of copy needed. This allows for best performances while sharing the same core for implementing memcpy(), copy_from_user() and copy_to_user() for instance. Two reasons for this work: 1) the current copy_to_user/copy_from_user implementation assumes no task switch will ever occur in the middle of each copied page making it completely unsafe with CONFIG_PREEMPT=y. 2) current copy implementations are measurably suboptimal and optimizing different implementations separately is a pain and more opportunities for bugs. The reason for (1) is the fact that copy inside user pages are performed with the ldm instruction which has no mean for testing user protections and could possibly race with process preemption bypassing the COW mechanism for example. This is a longstanding issue that we said ought to be fixed for about two years now. The solution is to substitute those ldm insns with a series of ldrt or strt insns to enforce user memory protection. At least on StrongARM and XScale cores the ldm is not faster than the equivalent ldr/str insns with a warm i-cache so there is no measurable performance degradation with that change. The fact that the copy code is a template makes it pretty easy to reuse the same core code as for memcpy and benefit from the same performance optimizations. Now (2) is best demonstrated with actual throughput measurements. First, here is a summary of memcopy tests performed on a StrongARM core: PTR alignment buffer size kernel version this version ------------------------------------------------------------ aligned 32 59.73 107.43 unaligned 32 61.31 74.72 aligned 100 132.47 136.15 unaligned 100 103.84 123.76 aligned 4096 130.67 130.80 unaligned 4096 130.68 130.64 aligned 1048576 68.03 68.18 unaligned 1048576 68.03 68.18 The buffer size is in bytes and the measured speed in MB/s. The copy was performed repeatedly with given buffer and throughput averaged over 3 seconds. Here we can see that the current kernel version has a higher entry cost that shows up with small buffers. As buffer size grows both implementation converge to the same throughput. Now here's the exact same test performed on an XScale core (PXA255): PTR alignment buffer size kernel version this version ------------------------------------------------------------ aligned 32 46.99 77.58 unaligned 32 53.61 59.59 aligned 100 107.19 136.59 unaligned 100 83.61 97.58 aligned 4096 129.13 129.98 unaligned 4096 128.36 128.53 aligned 1048576 53.76 59.41 unaligned 1048576 33.67 56.96 Again we can see the entry setup cost being higher for the current kernel before getting to the main copy loop. Then throughput results converge as long as the buffer remains in the cache. Then the 1MB case shows more differences probably due to better pld placement and/or less instruction interlocks in this proposed implementation. Disclaimer: The PXA system was running with slower clocks than the StrongARM system so trying to infer any conclusion by comparing those separate sets of results side by side would be completely inappropriate. So... What this patch does is to replace both memcpy and memmove with an implementation based on the provided copy code template. The memmove code is kept separate since it is used only if the memory areas involved do overlap in which case the code is a transposition of the template but with the copy occurring in the opposite direction (trying to fit that mode into the template turned it into a mess not worth it for memmove alone). And obviously both memcpy and memmove were tested with all kinds of pointer alignments and buffer sizes to exercise all code paths for correctness. The next patch will provide the now trivial replacement implementation copy_to_user and copy_from_user. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2005-11-01 12:52:23 -07:00
/*
* linux/arch/arm/lib/copy_template.s
*
* Code template for optimized memory copy functions
*
* Author: Nicolas Pitre
* Created: Sep 28, 2005
* Copyright: MontaVista Software, Inc.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as
* published by the Free Software Foundation.
*/
/*
* This can be used to enable code to cacheline align the source pointer.
* Experiments on tested architectures (StrongARM and XScale) didn't show
* this a worthwhile thing to do. That might be different in the future.
*/
//#define CALGN(code...) code
#define CALGN(code...)
/*
* Theory of operation
* -------------------
*
* This file provides the core code for a forward memory copy used in
* the implementation of memcopy(), copy_to_user() and copy_from_user().
*
* The including file must define the following accessor macros
* according to the need of the given function:
*
* ldr1w ptr reg abort
*
* This loads one word from 'ptr', stores it in 'reg' and increments
* 'ptr' to the next word. The 'abort' argument is used for fixup tables.
*
* ldr4w ptr reg1 reg2 reg3 reg4 abort
* ldr8w ptr, reg1 reg2 reg3 reg4 reg5 reg6 reg7 reg8 abort
*
* This loads four or eight words starting from 'ptr', stores them
* in provided registers and increments 'ptr' past those words.
* The'abort' argument is used for fixup tables.
*
* ldr1b ptr reg cond abort
*
* Similar to ldr1w, but it loads a byte and increments 'ptr' one byte.
* It also must apply the condition code if provided, otherwise the
* "al" condition is assumed by default.
*
* str1w ptr reg abort
* str8w ptr reg1 reg2 reg3 reg4 reg5 reg6 reg7 reg8 abort
* str1b ptr reg cond abort
*
* Same as their ldr* counterparts, but data is stored to 'ptr' location
* rather than being loaded.
*
* enter reg1 reg2
*
* Preserve the provided registers on the stack plus any additional
* data as needed by the implementation including this code. Called
* upon code entry.
*
* exit reg1 reg2
*
* Restore registers with the values previously saved with the
* 'preserv' macro. Called upon code termination.
*/
enter r4, lr
subs r2, r2, #4
blt 8f
ands ip, r0, #3
PLD( pld [r1, #0] )
bne 9f
ands ip, r1, #3
bne 10f
1: subs r2, r2, #(28)
stmfd sp!, {r5 - r8}
blt 5f
CALGN( ands ip, r1, #31 )
CALGN( rsb r3, ip, #32 )
CALGN( sbcnes r4, r3, r2 ) @ C is always set here
CALGN( bcs 2f )
CALGN( adr r4, 6f )
CALGN( subs r2, r2, r3 ) @ C gets set
CALGN( add pc, r4, ip )
PLD( pld [r1, #0] )
2: PLD( subs r2, r2, #96 )
PLD( pld [r1, #28] )
PLD( blt 4f )
PLD( pld [r1, #60] )
PLD( pld [r1, #92] )
3: PLD( pld [r1, #124] )
4: ldr8w r1, r3, r4, r5, r6, r7, r8, ip, lr, abort=20f
subs r2, r2, #32
str8w r0, r3, r4, r5, r6, r7, r8, ip, lr, abort=20f
bge 3b
PLD( cmn r2, #96 )
PLD( bge 4b )
5: ands ip, r2, #28
rsb ip, ip, #32
addne pc, pc, ip @ C is always clear here
b 7f
6: nop
ldr1w r1, r3, abort=20f
ldr1w r1, r4, abort=20f
ldr1w r1, r5, abort=20f
ldr1w r1, r6, abort=20f
ldr1w r1, r7, abort=20f
ldr1w r1, r8, abort=20f
ldr1w r1, lr, abort=20f
add pc, pc, ip
nop
nop
str1w r0, r3, abort=20f
str1w r0, r4, abort=20f
str1w r0, r5, abort=20f
str1w r0, r6, abort=20f
str1w r0, r7, abort=20f
str1w r0, r8, abort=20f
str1w r0, lr, abort=20f
CALGN( bcs 2b )
7: ldmfd sp!, {r5 - r8}
8: movs r2, r2, lsl #31
ldr1b r1, r3, ne, abort=21f
ldr1b r1, r4, cs, abort=21f
ldr1b r1, ip, cs, abort=21f
str1b r0, r3, ne, abort=21f
str1b r0, r4, cs, abort=21f
str1b r0, ip, cs, abort=21f
exit r4, pc
9: rsb ip, ip, #4
cmp ip, #2
ldr1b r1, r3, gt, abort=21f
ldr1b r1, r4, ge, abort=21f
ldr1b r1, lr, abort=21f
str1b r0, r3, gt, abort=21f
str1b r0, r4, ge, abort=21f
subs r2, r2, ip
str1b r0, lr, abort=21f
blt 8b
ands ip, r1, #3
beq 1b
10: bic r1, r1, #3
cmp ip, #2
ldr1w r1, lr, abort=21f
beq 17f
bgt 18f
.macro forward_copy_shift pull push
subs r2, r2, #28
blt 14f
CALGN( ands ip, r1, #31 )
CALGN( rsb ip, ip, #32 )
CALGN( sbcnes r4, ip, r2 ) @ C is always set here
CALGN( subcc r2, r2, ip )
CALGN( bcc 15f )
11: stmfd sp!, {r5 - r9}
PLD( pld [r1, #0] )
PLD( subs r2, r2, #96 )
PLD( pld [r1, #28] )
PLD( blt 13f )
PLD( pld [r1, #60] )
PLD( pld [r1, #92] )
12: PLD( pld [r1, #124] )
13: ldr4w r1, r4, r5, r6, r7, abort=19f
mov r3, lr, pull #\pull
subs r2, r2, #32
ldr4w r1, r8, r9, ip, lr, abort=19f
orr r3, r3, r4, push #\push
mov r4, r4, pull #\pull
orr r4, r4, r5, push #\push
mov r5, r5, pull #\pull
orr r5, r5, r6, push #\push
mov r6, r6, pull #\pull
orr r6, r6, r7, push #\push
mov r7, r7, pull #\pull
orr r7, r7, r8, push #\push
mov r8, r8, pull #\pull
orr r8, r8, r9, push #\push
mov r9, r9, pull #\pull
orr r9, r9, ip, push #\push
mov ip, ip, pull #\pull
orr ip, ip, lr, push #\push
str8w r0, r3, r4, r5, r6, r7, r8, r9, ip, , abort=19f
bge 12b
PLD( cmn r2, #96 )
PLD( bge 13b )
ldmfd sp!, {r5 - r9}
14: ands ip, r2, #28
beq 16f
15: mov r3, lr, pull #\pull
ldr1w r1, lr, abort=21f
subs ip, ip, #4
orr r3, r3, lr, push #\push
str1w r0, r3, abort=21f
bgt 15b
CALGN( cmp r2, #0 )
CALGN( bge 11b )
16: sub r1, r1, #(\push / 8)
b 8b
.endm
forward_copy_shift pull=8 push=24
17: forward_copy_shift pull=16 push=16
18: forward_copy_shift pull=24 push=8
/*
* Abort preamble and completion macros.
[ARM] 2947/1: copy template with new memcpy/memmove Patch from Nicolas Pitre This patch provides a new implementation for optimized memory copy functions on ARM. It is made of two levels: a template that consists of the core copy code and separate files that define macros to be used with the core code depending on the type of copy needed. This allows for best performances while sharing the same core for implementing memcpy(), copy_from_user() and copy_to_user() for instance. Two reasons for this work: 1) the current copy_to_user/copy_from_user implementation assumes no task switch will ever occur in the middle of each copied page making it completely unsafe with CONFIG_PREEMPT=y. 2) current copy implementations are measurably suboptimal and optimizing different implementations separately is a pain and more opportunities for bugs. The reason for (1) is the fact that copy inside user pages are performed with the ldm instruction which has no mean for testing user protections and could possibly race with process preemption bypassing the COW mechanism for example. This is a longstanding issue that we said ought to be fixed for about two years now. The solution is to substitute those ldm insns with a series of ldrt or strt insns to enforce user memory protection. At least on StrongARM and XScale cores the ldm is not faster than the equivalent ldr/str insns with a warm i-cache so there is no measurable performance degradation with that change. The fact that the copy code is a template makes it pretty easy to reuse the same core code as for memcpy and benefit from the same performance optimizations. Now (2) is best demonstrated with actual throughput measurements. First, here is a summary of memcopy tests performed on a StrongARM core: PTR alignment buffer size kernel version this version ------------------------------------------------------------ aligned 32 59.73 107.43 unaligned 32 61.31 74.72 aligned 100 132.47 136.15 unaligned 100 103.84 123.76 aligned 4096 130.67 130.80 unaligned 4096 130.68 130.64 aligned 1048576 68.03 68.18 unaligned 1048576 68.03 68.18 The buffer size is in bytes and the measured speed in MB/s. The copy was performed repeatedly with given buffer and throughput averaged over 3 seconds. Here we can see that the current kernel version has a higher entry cost that shows up with small buffers. As buffer size grows both implementation converge to the same throughput. Now here's the exact same test performed on an XScale core (PXA255): PTR alignment buffer size kernel version this version ------------------------------------------------------------ aligned 32 46.99 77.58 unaligned 32 53.61 59.59 aligned 100 107.19 136.59 unaligned 100 83.61 97.58 aligned 4096 129.13 129.98 unaligned 4096 128.36 128.53 aligned 1048576 53.76 59.41 unaligned 1048576 33.67 56.96 Again we can see the entry setup cost being higher for the current kernel before getting to the main copy loop. Then throughput results converge as long as the buffer remains in the cache. Then the 1MB case shows more differences probably due to better pld placement and/or less instruction interlocks in this proposed implementation. Disclaimer: The PXA system was running with slower clocks than the StrongARM system so trying to infer any conclusion by comparing those separate sets of results side by side would be completely inappropriate. So... What this patch does is to replace both memcpy and memmove with an implementation based on the provided copy code template. The memmove code is kept separate since it is used only if the memory areas involved do overlap in which case the code is a transposition of the template but with the copy occurring in the opposite direction (trying to fit that mode into the template turned it into a mess not worth it for memmove alone). And obviously both memcpy and memmove were tested with all kinds of pointer alignments and buffer sizes to exercise all code paths for correctness. The next patch will provide the now trivial replacement implementation copy_to_user and copy_from_user. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2005-11-01 12:52:23 -07:00
* If a fixup handler is required then those macros must surround it.
* It is assumed that the fixup code will handle the private part of
* the exit macro.
*/
.macro copy_abort_preamble
19: ldmfd sp!, {r5 - r9}
b 21f
20: ldmfd sp!, {r5 - r8}
21:
.endm
.macro copy_abort_end
ldmfd sp!, {r4, pc}
.endm