| Message ID | 20191008194552.2176475-1-paul.burton@mips.com (mailing list archive) |
|---|---|
| State | Mainlined |
| Commit | 3c0be5849259b729580c23549330973a2dd513a2 |
| Delegated to: | Paul Burton |
| Headers | show |
| Series | MIPS: Drop 32-bit asm string functions | expand |
On 10/8/19 9:46 PM, Paul Burton wrote: > We have assembly implementations of strcpy(), strncpy(), strcmp() & > strncmp() which: > > - Are simple byte-at-a-time loops with no particular optimizations. As > a comment in the code describes, they're "rather naive". > > - Offer no clear performance advantage over the generic C > implementations - in microbenchmarks performed by Alexander Lobakin > the asm functions sometimes win & sometimes lose, but generally not > by large margins in either direction. Maybe add: "See URL below for full benchmark results." > > - Don't support 64-bit kernels, where we already make use of the > generic C implementations. > > - Tend to bloat kernel code size due to inlining. > > - Don't support CONFIG_FORTIFY_SOURCE. > > - Won't support nanoMIPS without rework. > > For all of these reasons, delete the asm implementations & make use of > the generic C implementations for 32-bit kernels just like we already do > for 64-bit kernels. > > Signed-off-by: Paul Burton <paul.burton@mips.com> > URL: https://lore.kernel.org/linux-mips/a2a35f1cf58d6db19eb4af9b4ae21e35@dlink.ru/ > Cc: Alexander Lobakin <alobakin@dlink.ru> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> > --- > > arch/mips/include/asm/string.h | 123 --------------------------------- > 1 file changed, 123 deletions(-) > > diff --git a/arch/mips/include/asm/string.h b/arch/mips/include/asm/string.h > index 4b938c55b397..1de3bbce8e88 100644 > --- a/arch/mips/include/asm/string.h > +++ b/arch/mips/include/asm/string.h > @@ -10,129 +10,6 @@ > #ifndef _ASM_STRING_H > #define _ASM_STRING_H > > -#if !defined(__OPTIMIZE__) || !defined(CONFIG_FORTIFY_SOURCE) > - > -/* > - * Most of the inline functions are rather naive implementations so I just > - * didn't bother updating them for 64-bit ... > - */ > -#ifdef CONFIG_32BIT > - > -#ifndef IN_STRING_C > - > -#define __HAVE_ARCH_STRCPY > -static __inline__ char *strcpy(char *__dest, __const__ char *__src) > -{ > - char *__xdest = __dest; > - > - __asm__ __volatile__( > - ".set\tnoreorder\n\t" > - ".set\tnoat\n" > - "1:\tlbu\t$1,(%1)\n\t" > - "addiu\t%1,1\n\t" > - "sb\t$1,(%0)\n\t" > - "bnez\t$1,1b\n\t" > - "addiu\t%0,1\n\t" > - ".set\tat\n\t" > - ".set\treorder" > - : "=r" (__dest), "=r" (__src) > - : "0" (__dest), "1" (__src) > - : "memory"); > - > - return __xdest; > -} > - > -#define __HAVE_ARCH_STRNCPY > -static __inline__ char *strncpy(char *__dest, __const__ char *__src, size_t __n) > -{ > - char *__xdest = __dest; > - > - if (__n == 0) > - return __xdest; > - > - __asm__ __volatile__( > - ".set\tnoreorder\n\t" > - ".set\tnoat\n" > - "1:\tlbu\t$1,(%1)\n\t" > - "subu\t%2,1\n\t" > - "sb\t$1,(%0)\n\t" > - "beqz\t$1,2f\n\t" > - "addiu\t%0,1\n\t" > - "bnez\t%2,1b\n\t" > - "addiu\t%1,1\n" > - "2:\n\t" > - ".set\tat\n\t" > - ".set\treorder" > - : "=r" (__dest), "=r" (__src), "=r" (__n) > - : "0" (__dest), "1" (__src), "2" (__n) > - : "memory"); > - > - return __xdest; > -} > - > -#define __HAVE_ARCH_STRCMP > -static __inline__ int strcmp(__const__ char *__cs, __const__ char *__ct) > -{ > - int __res; > - > - __asm__ __volatile__( > - ".set\tnoreorder\n\t" > - ".set\tnoat\n\t" > - "lbu\t%2,(%0)\n" > - "1:\tlbu\t$1,(%1)\n\t" > - "addiu\t%0,1\n\t" > - "bne\t$1,%2,2f\n\t" > - "addiu\t%1,1\n\t" > - "bnez\t%2,1b\n\t" > - "lbu\t%2,(%0)\n\t" > -#if defined(CONFIG_CPU_R3000) > - "nop\n\t" > -#endif > - "move\t%2,$1\n" > - "2:\tsubu\t%2,$1\n" > - "3:\t.set\tat\n\t" > - ".set\treorder" > - : "=r" (__cs), "=r" (__ct), "=r" (__res) > - : "0" (__cs), "1" (__ct)); > - > - return __res; > -} > - > -#endif /* !defined(IN_STRING_C) */ > - > -#define __HAVE_ARCH_STRNCMP > -static __inline__ int > -strncmp(__const__ char *__cs, __const__ char *__ct, size_t __count) > -{ > - int __res; > - > - __asm__ __volatile__( > - ".set\tnoreorder\n\t" > - ".set\tnoat\n" > - "1:\tlbu\t%3,(%0)\n\t" > - "beqz\t%2,2f\n\t" > - "lbu\t$1,(%1)\n\t" > - "subu\t%2,1\n\t" > - "bne\t$1,%3,3f\n\t" > - "addiu\t%0,1\n\t" > - "bnez\t%3,1b\n\t" > - "addiu\t%1,1\n" > - "2:\n\t" > -#if defined(CONFIG_CPU_R3000) > - "nop\n\t" > -#endif > - "move\t%3,$1\n" > - "3:\tsubu\t%3,$1\n\t" > - ".set\tat\n\t" > - ".set\treorder" > - : "=r" (__cs), "=r" (__ct), "=r" (__count), "=r" (__res) > - : "0" (__cs), "1" (__ct), "2" (__count)); > - > - return __res; > -} > -#endif /* CONFIG_32BIT */ > -#endif /* !defined(__OPTIMIZE__) || !defined(CONFIG_FORTIFY_SOURCE) */ > - > #define __HAVE_ARCH_MEMSET > extern void *memset(void *__s, int __c, size_t __count); > >
Hello, Paul Burton wrote: > We have assembly implementations of strcpy(), strncpy(), strcmp() & > strncmp() which: > > - Are simple byte-at-a-time loops with no particular optimizations. As > a comment in the code describes, they're "rather naive". > > - Offer no clear performance advantage over the generic C > implementations - in microbenchmarks performed by Alexander Lobakin > the asm functions sometimes win & sometimes lose, but generally not > by large margins in either direction. > > - Don't support 64-bit kernels, where we already make use of the > generic C implementations. > > - Tend to bloat kernel code size due to inlining. > > - Don't support CONFIG_FORTIFY_SOURCE. > > - Won't support nanoMIPS without rework. > > For all of these reasons, delete the asm implementations & make use of > the generic C implementations for 32-bit kernels just like we already do > for 64-bit kernels. Applied to mips-next. > commit 3c0be5849259 > https://git.kernel.org/mips/c/3c0be5849259 > > Signed-off-by: Paul Burton <paul.burton@mips.com> > URL: https://lore.kernel.org/linux-mips/a2a35f1cf58d6db19eb4af9b4ae21e35@dlink.ru/ > Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Thanks, Paul [ This message was auto-generated; if you believe anything is incorrect then please email paul.burton@mips.com to report it. ]
diff --git a/arch/mips/include/asm/string.h b/arch/mips/include/asm/string.h index 4b938c55b397..1de3bbce8e88 100644 --- a/arch/mips/include/asm/string.h +++ b/arch/mips/include/asm/string.h @@ -10,129 +10,6 @@ #ifndef _ASM_STRING_H #define _ASM_STRING_H -#if !defined(__OPTIMIZE__) || !defined(CONFIG_FORTIFY_SOURCE) - -/* - * Most of the inline functions are rather naive implementations so I just - * didn't bother updating them for 64-bit ... - */ -#ifdef CONFIG_32BIT - -#ifndef IN_STRING_C - -#define __HAVE_ARCH_STRCPY -static __inline__ char *strcpy(char *__dest, __const__ char *__src) -{ - char *__xdest = __dest; - - __asm__ __volatile__( - ".set\tnoreorder\n\t" - ".set\tnoat\n" - "1:\tlbu\t$1,(%1)\n\t" - "addiu\t%1,1\n\t" - "sb\t$1,(%0)\n\t" - "bnez\t$1,1b\n\t" - "addiu\t%0,1\n\t" - ".set\tat\n\t" - ".set\treorder" - : "=r" (__dest), "=r" (__src) - : "0" (__dest), "1" (__src) - : "memory"); - - return __xdest; -} - -#define __HAVE_ARCH_STRNCPY -static __inline__ char *strncpy(char *__dest, __const__ char *__src, size_t __n) -{ - char *__xdest = __dest; - - if (__n == 0) - return __xdest; - - __asm__ __volatile__( - ".set\tnoreorder\n\t" - ".set\tnoat\n" - "1:\tlbu\t$1,(%1)\n\t" - "subu\t%2,1\n\t" - "sb\t$1,(%0)\n\t" - "beqz\t$1,2f\n\t" - "addiu\t%0,1\n\t" - "bnez\t%2,1b\n\t" - "addiu\t%1,1\n" - "2:\n\t" - ".set\tat\n\t" - ".set\treorder" - : "=r" (__dest), "=r" (__src), "=r" (__n) - : "0" (__dest), "1" (__src), "2" (__n) - : "memory"); - - return __xdest; -} - -#define __HAVE_ARCH_STRCMP -static __inline__ int strcmp(__const__ char *__cs, __const__ char *__ct) -{ - int __res; - - __asm__ __volatile__( - ".set\tnoreorder\n\t" - ".set\tnoat\n\t" - "lbu\t%2,(%0)\n" - "1:\tlbu\t$1,(%1)\n\t" - "addiu\t%0,1\n\t" - "bne\t$1,%2,2f\n\t" - "addiu\t%1,1\n\t" - "bnez\t%2,1b\n\t" - "lbu\t%2,(%0)\n\t" -#if defined(CONFIG_CPU_R3000) - "nop\n\t" -#endif - "move\t%2,$1\n" - "2:\tsubu\t%2,$1\n" - "3:\t.set\tat\n\t" - ".set\treorder" - : "=r" (__cs), "=r" (__ct), "=r" (__res) - : "0" (__cs), "1" (__ct)); - - return __res; -} - -#endif /* !defined(IN_STRING_C) */ - -#define __HAVE_ARCH_STRNCMP -static __inline__ int -strncmp(__const__ char *__cs, __const__ char *__ct, size_t __count) -{ - int __res; - - __asm__ __volatile__( - ".set\tnoreorder\n\t" - ".set\tnoat\n" - "1:\tlbu\t%3,(%0)\n\t" - "beqz\t%2,2f\n\t" - "lbu\t$1,(%1)\n\t" - "subu\t%2,1\n\t" - "bne\t$1,%3,3f\n\t" - "addiu\t%0,1\n\t" - "bnez\t%3,1b\n\t" - "addiu\t%1,1\n" - "2:\n\t" -#if defined(CONFIG_CPU_R3000) - "nop\n\t" -#endif - "move\t%3,$1\n" - "3:\tsubu\t%3,$1\n\t" - ".set\tat\n\t" - ".set\treorder" - : "=r" (__cs), "=r" (__ct), "=r" (__count), "=r" (__res) - : "0" (__cs), "1" (__ct), "2" (__count)); - - return __res; -} -#endif /* CONFIG_32BIT */ -#endif /* !defined(__OPTIMIZE__) || !defined(CONFIG_FORTIFY_SOURCE) */ - #define __HAVE_ARCH_MEMSET extern void *memset(void *__s, int __c, size_t __count);
We have assembly implementations of strcpy(), strncpy(), strcmp() & strncmp() which: - Are simple byte-at-a-time loops with no particular optimizations. As a comment in the code describes, they're "rather naive". - Offer no clear performance advantage over the generic C implementations - in microbenchmarks performed by Alexander Lobakin the asm functions sometimes win & sometimes lose, but generally not by large margins in either direction. - Don't support 64-bit kernels, where we already make use of the generic C implementations. - Tend to bloat kernel code size due to inlining. - Don't support CONFIG_FORTIFY_SOURCE. - Won't support nanoMIPS without rework. For all of these reasons, delete the asm implementations & make use of the generic C implementations for 32-bit kernels just like we already do for 64-bit kernels. Signed-off-by: Paul Burton <paul.burton@mips.com> URL: https://lore.kernel.org/linux-mips/a2a35f1cf58d6db19eb4af9b4ae21e35@dlink.ru/ Cc: Alexander Lobakin <alobakin@dlink.ru> --- arch/mips/include/asm/string.h | 123 --------------------------------- 1 file changed, 123 deletions(-)