target/linux/generic/pending-4.14/306-mips_mem_functions_performance.patch

   1 From: Felix Fietkau <nbd@nbd.name>
   2 Subject: [PATCH] mips: allow the compiler to optimize memset, memcmp, memcpy  for better performance and (in some instances) smaller code
   3
   4 lede-commit: 07e59c7bc7f375f792ec9734be42fe4fa391a8bb
   5 Signed-off-by: Felix Fietkau <nbd@nbd.name>
   6 ---
   7  arch/mips/boot/compressed/Makefile |  3 ++-
   8  arch/mips/include/asm/string.h     | 38 ++++++++++++++++++++++++++++++++++++++
   9  arch/mips/lib/Makefile             |  2 +-
  10  arch/mips/lib/memcmp.c             | 22 ++++++++++++++++++++++
  11  4 files changed, 63 insertions(+), 2 deletions(-)
  12  create mode 100644 arch/mips/lib/memcmp.c
  13
  14 --- a/arch/mips/boot/compressed/Makefile
  15 +++ b/arch/mips/boot/compressed/Makefile
  16 @@ -23,7 +23,8 @@ KBUILD_CFLAGS := $(filter-out -pg, $(KBU
  17  KBUILD_CFLAGS := $(filter-out -fstack-protector, $(KBUILD_CFLAGS))
  18
  19  KBUILD_CFLAGS := $(KBUILD_CFLAGS) -D__KERNEL__ \
  20 -       -DBOOT_HEAP_SIZE=$(BOOT_HEAP_SIZE) -D"VMLINUX_LOAD_ADDRESS_ULL=$(VMLINUX_LOAD_ADDRESS)ull"
  21 +       -DBOOT_HEAP_SIZE=$(BOOT_HEAP_SIZE) -D"VMLINUX_LOAD_ADDRESS_ULL=$(VMLINUX_LOAD_ADDRESS)ull" \
  22 +       -D__ZBOOT__
  23
  24  KBUILD_AFLAGS := $(KBUILD_AFLAGS) -D__ASSEMBLY__ \
  25         -DBOOT_HEAP_SIZE=$(BOOT_HEAP_SIZE) \
  26 --- a/arch/mips/include/asm/string.h
  27 +++ b/arch/mips/include/asm/string.h
  28 @@ -140,4 +140,42 @@ extern void *memcpy(void *__to, __const_
  29  #define __HAVE_ARCH_MEMMOVE
  30  extern void *memmove(void *__dest, __const__ void *__src, size_t __n);
  31
  32 +#ifndef __ZBOOT__
  33 +#define memset(__s, __c, len)                                  \
  34 +({                                                             \
  35 +       size_t __len = (len);                                   \
  36 +       void *__ret;                                            \
  37 +       if (__builtin_constant_p(len) && __len >= 64)           \
  38 +               __ret = memset((__s), (__c), __len);            \
  39 +       else                                                    \
  40 +               __ret = __builtin_memset((__s), (__c), __len);  \
  41 +       __ret;                                                  \
  42 +})
  43 +
  44 +#define memcpy(dst, src, len)                                  \
  45 +({                                                             \
  46 +       size_t __len = (len);                                   \
  47 +       void *__ret;                                            \
  48 +       if (__builtin_constant_p(len) && __len >= 64)           \
  49 +               __ret = memcpy((dst), (src), __len);            \
  50 +       else                                                    \
  51 +               __ret = __builtin_memcpy((dst), (src), __len);  \
  52 +       __ret;                                                  \
  53 +})
  54 +
  55 +#define memmove(dst, src, len)                                 \
  56 +({                                                             \
  57 +       size_t __len = (len);                                   \
  58 +       void *__ret;                                            \
  59 +       if (__builtin_constant_p(len) && __len >= 64)           \
  60 +               __ret = memmove((dst), (src), __len);           \
  61 +       else                                                    \
  62 +               __ret = __builtin_memmove((dst), (src), __len); \
  63 +       __ret;                                                  \
  64 +})
  65 +
  66 +#define __HAVE_ARCH_MEMCMP
  67 +#define memcmp(src1, src2, len) __builtin_memcmp((src1), (src2), (len))
  68 +#endif
  69 +
  70  #endif /* _ASM_STRING_H */
  71 --- a/arch/mips/lib/Makefile
  72 +++ b/arch/mips/lib/Makefile
  73 @@ -5,7 +5,7 @@
  74
  75  lib-y  += bitops.o csum_partial.o delay.o memcpy.o memset.o \
  76            mips-atomic.o strncpy_user.o \
  77 -          strnlen_user.o uncached.o
  78 +          strnlen_user.o uncached.o memcmp.o
  79
  80  obj-y                  += iomap.o iomap_copy.o
  81  obj-$(CONFIG_PCI)      += iomap-pci.o
  82 --- /dev/null
  83 +++ b/arch/mips/lib/memcmp.c
  84 @@ -0,0 +1,22 @@
  85 +/*
  86 + *  copied from linux/lib/string.c
  87 + *
  88 + *  Copyright (C) 1991, 1992  Linus Torvalds
  89 + */
  90 +
  91 +#include <linux/module.h>
  92 +#include <linux/string.h>
  93 +
  94 +#undef memcmp
  95 +int memcmp(const void *cs, const void *ct, size_t count)
  96 +{
  97 +       const unsigned char *su1, *su2;
  98 +       int res = 0;
  99 +
 100 +       for (su1 = cs, su2 = ct; 0 < count; ++su1, ++su2, count--)
 101 +               if ((res = *su1 - *su2) != 0)
 102 +                       break;
 103 +       return res;
 104 +}
 105 +EXPORT_SYMBOL(memcmp);
 106 +