32-bit MMIO copy routines, reworked
Subsystems (11)
PATCH 1/1 32-bit MMIO copy routines, reworked
After some more review comments from Roland, Andrew and Chris Hellwig, here is a reworked set of 32-bit MMIO copy patches.
These use CONFIG_RAW_MEMCPY_IO to determine whether an arch should use the generic __raw_memcpy_toio32 routine or its own specialised version. We provide a specialised implementation for x86_64.
These patches should apply cleanly against current -git, and have been tested on i386 and x86_64.
The patch series is as follows:
raw_memcpy_io.patch
Introduce the generic MMIO 32-bit copy routine.
x86_64-memcpy32.patch
Add memcpy32 routine to x86_64.
arch-specific-raw_memcpy_io.patch
Get each arch to use generic memcpy_io code, except x86_64, which uses memcpy32.
PATCH 1/1 Introduce __raw_memcpy_toio32
This arch-independent routine copies data to a memory-mapped I/O region, using 32-bit accesses. It does not guarantee access ordering, nor does it perform a memory barrier afterwards. This style of access is required by some devices.
lib/Makefile +1 -0
@@ -21,6 +21,7 @@ lib-$(CONFIG_RWSEM_XCHGADD_ALGORITHM) += rwsem.o lib-$(CONFIG_SEMAPHORE_SLEEPERS) += semaphore-sleepers.o lib-$(CONFIG_GENERIC_FIND_NEXT_BIT) += find_next_bit.o +lib-$(CONFIG_GENERIC_RAW_MEMCPY_IO) += raw_memcpy_io.o obj-$(CONFIG_LOCK_KERNEL) += kernel_lock.o obj-$(CONFIG_DEBUG_PREEMPT) += smp_processor_id.o
lib/raw_memcpy_io.c new file +39 -0
@@ -0,0 +1,39 @@ +/* +* Copyright 2006 PathScale, Inc. All Rights Reserved. +* +* This file is free software; you can redistribute it and/or modify +* it under the terms of version 2 of the GNU General Public License +* as published by the Free Software Foundation. +* +* This program 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 this program; if not, write to the Free Software Foundation, +* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. +*/ + +#include <linux/types.h> +#include <asm/io.h> + +/** +* __raw_memcpy_toio32 - copy data to MMIO space, in 32-bit units +* @to: destination, in MMIO space (must be 32-bit aligned) +* @from: source (must be 32-bit aligned) +* @count: number of 32-bit quantities to copy +* +* Copy data from kernel space to MMIO space, in units of 32 bits at a +* time. Order of access is not guaranteed, nor is a memory barrier +* performed afterwards. +*/ +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count) +{ +u32 __iomem *dst = to; +const u32 *src = from; +size_t i; + +for (i =0 ; i < count; i++) +__raw_writel (*src++, dst++); +}
PATCH 1/1 Add __raw_memcpy_toio32 to each arch
In reply to Bryan O'SullivanMost arches use the generic routine. x86_64 uses memcpy32 instead; this is substantially faster, even over a bus that is much slower than the CPU.
arch/alpha/Kconfig +4 -0
@@ -42,6 +42,10 @@ default y config GENERIC_IRQ_PROBE + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/arm/Kconfig +4 -0
@@ -59,6 +59,10 @@ config GENERIC_BUST_SPINLOCK bool + +config GENERIC_RAW_MEMCPY_IO + bool + default y config ARCH_MAY_HAVE_PC_FDC bool
arch/arm26/Kconfig +4 -0
@@ -33,6 +33,10 @@ config FORCE_MAX_ZONEORDER int default 9 + +config GENERIC_RAW_MEMCPY_IO + bool + default y config RWSEM_GENERIC_SPINLOCK bool
arch/cris/Kconfig +4 -0
@@ -17,6 +17,10 @@ bool config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/frv/Kconfig +4 -0
@@ -24,6 +24,10 @@ config GENERIC_CALIBRATE_DELAY bool default n + +config GENERIC_RAW_MEMCPY_IO + bool + default y config GENERIC_HARDIRQS bool
arch/h8300/Kconfig +4 -0
@@ -30,6 +30,10 @@ default n config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/i386/Kconfig +4 -0
@@ -34,6 +34,10 @@ default y config GENERIC_IOMAP + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/ia64/Kconfig +4 -0
@@ -47,6 +47,10 @@ default y config GENERIC_IOMAP + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/m32r/Kconfig +4 -0
@@ -25,6 +25,10 @@ default y config GENERIC_IRQ_PROBE + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/m68k/Kconfig +4 -0
@@ -18,6 +18,10 @@ bool config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/m68knommu/Kconfig +4 -0
@@ -26,6 +26,10 @@ default n config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/mips/Kconfig +4 -0
@@ -800,6 +800,10 @@ bool config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/parisc/Kconfig +4 -0
@@ -37,6 +37,10 @@ config GENERIC_IRQ_PROBE def_bool y + +config GENERIC_RAW_MEMCPY_IO + bool + default y # unless you want to implement ACPI on PA-RISC ... ;-) config PM
arch/powerpc/Kconfig +4 -0
@@ -38,6 +38,10 @@ default y config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/ppc/Kconfig +4 -0
@@ -16,6 +16,10 @@ bool config RWSEM_XCHGADD_ALGORITHM + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/s390/Kconfig +4 -0
@@ -20,6 +20,10 @@ config GENERIC_BUST_SPINLOCK bool + +config GENERIC_RAW_MEMCPY_IO + bool + default y mainmenu "Linux Kernel Configuration"
arch/sh/Kconfig +4 -0
@@ -30,6 +30,10 @@ default y config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/sh64/Kconfig +4 -0
@@ -34,6 +34,10 @@ config GENERIC_ISA_DMA bool + +config GENERIC_RAW_MEMCPY_IO + bool + default y source init/Kconfig
arch/sparc/Kconfig +4 -0
@@ -152,6 +152,10 @@ bool config GENERIC_CALIBRATE_DELAY + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
arch/sparc64/Kconfig +4 -0
@@ -166,6 +166,10 @@ bool default y +config GENERIC_RAW_MEMCPY_IO + bool + default y + choice prompt "SPARC64 Huge TLB Page Size" depends on HUGETLB_PAGE
arch/v850/Kconfig +3 -0
@@ -17,6 +17,9 @@ bool default n config GENERIC_CALIBRATE_DELAY + bool + default y +config GENERIC_RAW_MEMCPY_IO bool default y
arch/xtensa/Kconfig +4 -0
@@ -27,6 +27,10 @@ default y config GENERIC_HARDIRQS + bool + default y + +config GENERIC_RAW_MEMCPY_IO bool default y
include/asm-alpha/io.h +2 -0
@@ -504,6 +504,8 @@extern void memcpy_toio (volatile void __iomem *,const void *,long );extern void _memset_c_io (volatile void __iomem *,unsigned long ,long ); +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +static inline void memset_io (volatile void __iomem *addr,u8 c,long len) {_memset_c_io (addr,0x0101010101010101UL * c, len);
include/asm-arm/io.h +2 -0
@@ -189,6 +189,8 @@#define memset_io (c,v,l) _memset_io(__mem_pci(c),(v),(l))#define memcpy_fromio (a,c,l) _memcpy_fromio((a),__mem_pci(c),(l))#define memcpy_toio (c,a,l) _memcpy_toio(__mem_pci(c),(a),(l)) + +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count);#define eth_io_copy_and_sum (s,c,l,b) \ eth_copy_and_sum((s),__mem_pci(c),(l),(b))
include/asm-cris/io.h +2 -0
@@ -121,6 +121,8 @@#define memcpy_fromio (a,b,c) memcpy((a),(void *)(b),(c))#define memcpy_toio (a,b,c) memcpy((void *)(a),(b),(c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* * Again, CRIS does not require mem IO specific function. */
include/asm-frv/io.h +2 -0
@@ -127,6 +127,8 @@memcpy ((void __force *) dst, src, count); } +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +static inline uint8_t inb (unsigned long addr) {return __builtin_read8 ((void *)addr);
include/asm-h8300/io.h +2 -0
@@ -209,6 +209,8 @@#define memcpy_fromio (a,b,c) memcpy((a),(void *)(b),(c))#define memcpy_toio (a,b,c) memcpy((void *)(a),(b),(c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define mmiowb ()#define inb (addr) ((h8300_buswidth(addr))?readw((addr) & ~1) & 0xff:readb(addr))
include/asm-i386/io.h +2 -0
@@ -203,6 +203,8 @@ {__memcpy ((void __force *) dst, src, count); } + +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count);/* * ISA space is 'always mapped' on a typical x86 system, no need to
include/asm-ia64/io.h +2 -0
@@ -444,6 +444,8 @@extern void memcpy_toio (volatile void __iomem *dst,const void *src,long n);extern void memset_io (volatile void __iomem *s,int c,long n); +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define dma_cache_inv (_start,_size) do { } while (0)#define dma_cache_wback (_start,_size) do { } while (0)#define dma_cache_wback_inv (_start,_size) do { } while (0)
include/asm-m32r/io.h +2 -0
@@ -216,6 +216,8 @@memcpy ((void __force *) dst, src, count); } +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* * Convert a physical pointer to a virtual kernel pointer for /dev/mem * access
include/asm-m68knommu/io.h +2 -0
@@ -113,6 +113,8 @@#define memcpy_fromio (a,b,c) memcpy((a),(void *)(b),(c))#define memcpy_toio (a,b,c) memcpy((void *)(a),(b),(c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define inb (addr) readb(addr)#define inw (addr) readw(addr)#define inl (addr) readl(addr)
include/asm-mips/io.h +2 -0
@@ -534,6 +534,8 @@memcpy ((void __force *) dst, src, count); } +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* * Memory Mapped I/O */
include/asm-parisc/io.h +2 -0
@@ -294,6 +294,8 @@void memcpy_fromio (void *dst,const volatile void __iomem *src,int count);void memcpy_toio (volatile void __iomem *dst,const void *src,int count); +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* Support old drivers which don't ioremap. * NB this interface is scheduled to disappear in 2.5 */
include/asm-powerpc/io.h +2 -0
@@ -64,6 +64,8 @@#define memcpy_fromio (a,b,c) iSeries_memcpy_fromio((a), (b), (c))#define memcpy_toio (a,b,c) iSeries_memcpy_toio((a), (b), (c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define inb (addr) readb(((void __iomem *)(long)(addr)))#define inw (addr) readw(((void __iomem *)(long)(addr)))#define inl (addr) readl(((void __iomem *)(long)(addr)))
include/asm-ppc/io.h +2 -0
@@ -369,6 +369,8 @@ }#endif +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define eth_io_copy_and_sum (a,b,c,d) eth_copy_and_sum((a),(void __force *)(void __iomem *)(b),(c),(d))/*
include/asm-s390/io.h +2 -0
@@ -99,6 +99,8 @@#define memcpy_fromio (a,b,c) memcpy((a),__io_virt(b),(c))#define memcpy_toio (a,b,c) memcpy(__io_virt(a),(b),(c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#define inb_p (addr) readb(addr)#define inb (addr) readb(addr)
include/asm-sh/io.h +2 -0
@@ -177,6 +177,8 @@extern void memcpy_toio (unsigned long ,const void *,unsigned long );extern void memset_io (unsigned long ,int ,unsigned long ); +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* SuperH on-chip I/O functions */ static __inline__unsigned char ctrl_inb (unsigned long addr) {
include/asm-sh64/io.h +2 -0
@@ -125,6 +125,8 @@void memcpy_toio (void __iomem *to,const void *from,long count);void memcpy_fromio (void *to,void __iomem *from,long count); + +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count);#define mmiowb ()
include/asm-sparc/io.h +2 -0
@@ -239,6 +239,8 @@#define memcpy_toio (d,s,sz) _memcpy_toio(d,s,sz) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +#ifdef __KERNEL__/*
include/asm-sparc64/io.h +2 -0
@@ -440,6 +440,8 @@#define memcpy_toio (d,s,sz) _memcpy_toio(d,s,sz) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +static inline int check_signature (void __iomem *io_addr,const unsigned char *signature,int length)
include/asm-v850/io.h +2 -0
@@ -130,6 +130,8 @@#define memcpy_fromio (dst, src, len) memcpy (dst, (void *)src, len)#define memcpy_toio (dst, src, len) memcpy ((void *)dst, src, len) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* * Convert a physical pointer to a virtual kernel pointer for /dev/mem * access
include/asm-x86_64/io.h +8 -0
@@ -252,6 +252,14 @@__memcpy_toio ((unsigned long )to,from,len); } +#include <asm/string.h> + +/* See lib/raw_memcpy_io.c for kernel doc. */ +static inline void __raw_memcpy_toio32 (void __iomem *dst,const void *src,size_t count) +{ +memcpy32 ((void __force *) dst, src, count); +} +void memset_io (volatile void __iomem *a,int b,size_t c);/*
include/asm-xtensa/io.h +2 -0
@@ -159,6 +159,8 @@#define memcpy_fromio (a,b,c) memcpy((a),(void *)(b),(c))#define memcpy_toio (a,b,c) memcpy((void *)(a),(b),(c)) +void __raw_memcpy_toio32 (void __iomem *to,const void *from,size_t count); +/* At this point the Xtensa doesn't provide byte swap instructions */ #ifdef __XTENSA_EB__
Re: [PATCH 3 of 3] Add __raw_memcpy_toio32 to each arch
Quoted 3 lines · Bryan O'Sullivan
Most arches use the generic routine.
x86_64uses memcpy32 instead; this is substantially faster, even over a bus that is much slower than the CPU.
So did you run numbers against the C implementation with -funroll-loops ? What were the results?
-Andi
Re: [PATCH 3 of 3] Add __raw_memcpy_toio32 to each arch
Quoted 7 lines · Andi Kleen
Quoted 3 lines · Bryan O'Sullivan
Most arches use the generic routine.
x86_64uses memcpy32 instead; this is substantially faster, even over a bus that is much slower than the CPU.So did you run numbers against the C implementation with -funroll-loops ? What were the results?
The C implementation is about 5% slower when copying over HyperTransport.
<b
PATCH 1/1 memcpy32 for x86_64
In reply to Bryan O'SullivanIntroduce an x86_64-specific memcpy32 routine. The routine is similar to memcpy, but is guaranteed to work in units of 32 bits at a time.
arch/x86_64/kernel/x8664_ksyms.c +2 -0
@@ -164,6 +164,8 @@EXPORT_SYMBOL (memcpy);EXPORT_SYMBOL (__memcpy); +EXPORT_SYMBOL_GPL (memcpy32); +#ifdef CONFIG_RWSEM_XCHGADD_ALGORITHM /* prototypes are wrong, these are assembly with custom calling functions */ extern void rwsem_down_read_failed_thunk (void );
arch/x86_64/lib/Makefile +1 -1
@@ -9,4 +9,4 @@ lib-y := csum-partial.o csum-copy.o csum-wrappers.o delay.o \ usercopy.o getuser.o putuser.o \ thunk.o clear_page.o copy_page.o bitstr.o bitops.o -lib-y += memcpy.o memmove.o memset.o copy_user.o +lib-y += memcpy.o memcpy32.o memmove.o memset.o copy_user.o
include/asm-x86_64/string.h +3 -0
@@ -45,6 +45,9 @@#define __HAVE_ARCH_MEMMOVEvoid *memmove (void * dest,const void *src,size_t count); +/* copy data, 32 bits at a time */ +void memcpy32 (void *dst,const void *src,size_t count); +/* Use C out of line version for memcmp */ #define memcmp __builtin_memcmpint memcmp (const void * cs,const void * ct,size_t count);
arch/x86_64/lib/memcpy32.S new file +39 -0
@@ -0,0 +1,39 @@ +/* + * Copyright 2006 PathScale, Inc. All Rights Reserved. + * + * This file is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License + * as published by the Free Software Foundation. + * + * This program 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 this program; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. + */ + +/* + * Registers used below: + * dst - rdi + * src - rsi + * count - rdx + */ + +/** + * memcpy32 - copy data, in units of 32 bits at a time + * @dst: destination (must be 32-bit aligned) + * @src: source (must be 32-bit aligned) + * @count: number of 32-bit quantities to copy + */ + .globl memcpy32 +memcpy32: + movl %edx,%ecx + shrl $1,%ecx + andl $1,%edx + rep movsq + movl %edx,%ecx + rep movsd + ret
Re: [PATCH 2 of 3] memcpy32 for x86_64
Quoted 72 lines · Bryan O'Sullivan
Introduce an x86_64-specific memcpy32 routine. The routine is similar to memcpy, but is guaranteed to work in units of 32 bits at a time.
Signed-off-by Bryan O'Sullivan <[email protected]>diff -r
2d4af213d9c5-rb4863171295farch/x86_64/kernel/x8664_ksyms.carch/x86_64/kernel/x8664_ksyms.c +2
-0@@ -164,6 +164,8 @@EXPORT_SYMBOL (memcpy);EXPORT_SYMBOL (__memcpy); +EXPORT_SYMBOL_GPL (memcpy32); +#ifdef CONFIG_RWSEM_XCHGADD_ALGORITHM /* prototypes are wrong, these are assembly with custom calling functions */ extern void rwsem_down_read_failed_thunk (void );diff -r
2d4af213d9c5-rb4863171295farch/x86_64/lib/Makefilearch/x86_64/lib/Makefile +1
-1@@ -9,4 +9,4 @@ lib-y := csum-partial.o csum-copy.o csum-wrappers.o delay.o \ usercopy.o getuser.o putuser.o \ thunk.o clear_page.o copy_page.o bitstr.o bitops.o -lib-y += memcpy.o memmove.o memset.o copy_user.o +lib-y += memcpy.o memcpy32.o memmove.o memset.o copy_user.odiff -r
2d4af213d9c5-rb4863171295finclude/asm-x86_64/string.hinclude/asm-x86_64/string.h +3
-0@@ -45,6 +45,9 @@#define __HAVE_ARCH_MEMMOVEvoid *memmove (void * dest,const void *src,size_t count); +/* copy data, 32 bits at a time */ +void memcpy32 (void *dst,const void *src,size_t count); +/* Use C out of line version for memcmp */ #define memcmp __builtin_memcmpint memcmp (const void * cs,const void * ct,size_t count);diff -r
2d4af213d9c5-rb4863171295farch/x86_64/lib/memcpy32.Sarch/x86_64/lib/memcpy32.S new file +29
-0@@ -0,0 +1,39 @@ +/* + * Copyright 2006 PathScale, Inc. All Rights Reserved. + * + * This file is free software; you can redistribute it and/or modify + * it under the terms of version 2 of the GNU General Public License + * as published by the Free Software Foundation. + * + * This program 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 this program; if not, write to the Free Software Foundation, + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. + */ + +/* + * Registers used below: + * dst - rdi + * src - rsi + * count - rdx + */ + +/** + * memcpy32 - copy data, in units of 32 bits at a time + * @dst: destination (must be 32-bit aligned) + * @src: source (must be 32-bit aligned) + * @count: number of 32-bit quantities to copy
arch/x86_64/lib/memcpy32.S line 39
Quoted 10 lines
arch/x86_64/lib/memcpy32.S new file +10
-0+ */ + .globl memcpy32 +memcpy32: + movl %edx,%ecx + shrl $1,%ecx + andl $1,%edx + rep movsq + movl %edx,%ecx + rep movsd + ret
movsq is not a 32bit move, it's a 64 bit one.
There are three possibilities here:
1) I misunderstand what memcpy32 means (I understand it like "it guarantees that all accesses will be strictly 32bit")
2) On all current x86_64 hardware each 64bit access from/to IO mapped addresses is always converted to two 32bit accesses.
3) code is buggy
If it is (1) or (2), consider adding a comment to clear future reader's confusion.
--
vda
Re: [PATCH 2 of 3] memcpy32 for x86_64
Quoted 2 lines · Denis Vlasenko
2) On all current
x86_64hardware each 64bit access from/to IO mapped addresses is always converted to two 32bit accesses.
This is true for 64-bit writes over Hypertransport (reads don't get split up this way), but not for PCI-Express memory writes, which remain atomic 64-bit. I'll be converting the 64-bit accesses to 32-bit, as you and Andi suggested.
<b
Re: [PATCH 2 of 3] memcpy32 for x86_64
Quoted 1 line · Bryan O'Sullivan
This is true for 64-bit writes over Hypertransport
is this something that will always be or just something current hardware does?
Re: [PATCH 2 of 3] memcpy32 for x86_64
Quoted 6 lines · Chris Wedgwood
Quoted 1 line · Bryan O'Sullivan
This is true for 64-bit writes over Hypertransport
is this something that will always be or just something current hardware does?
Yes, why risking that things will go wrong?
Also you'll get shorter code. Instead of
Quoted 9 lines
+ .globl memcpy32
+memcpy32:
+ movl %edx,%ecx
+ shrl $1,%ecx
+ andl $1,%edx
+ rep movsq
+ movl %edx,%ecx
+ rep movsd
+ ret
you need just
.globl memcpy32
memcpy32:
movl %edx,%ecx
rep movsd
ret
With properly written inlined asms code will be
reduced to just "rep movsd".
--
vda
Re: [PATCH 2 of 3] memcpy32 for x86_64
Quoted 7 lines · Denis Vlasenko
you need just
.globl memcpy32
memcpy32:
movl %edx,%ecx
rep movsd
ret
This is what the current version of the patches in -mm does.
<b