x86, mem: Optimize memmove for small size and unaligned cases
Ma Ling [Thu, 16 Sep 2010 19:12:40 +0000 (03:12 +0800)]
movs instruction will combine data to accelerate moving data,
however we need to concern two cases about it.

1. movs instruction need long lantency to startup,
   so here we use general mov instruction to copy data.
2. movs instruction is not good for unaligned case,
   even if src offset is 0x10, dest offset is 0x0,
   we avoid and handle the case by general mov instruction.

Signed-off-by: Ma Ling <ling.ma@intel.com>
LKML-Reference: <1284664360-6138-1-git-send-email-ling.ma@intel.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>

arch/x86/lib/memcpy_32.c
arch/x86/lib/memmove_64.c

index 81130d4..b908a59 100644 (file)
@@ -22,36 +22,187 @@ EXPORT_SYMBOL(memset);
 
 void *memmove(void *dest, const void *src, size_t n)
 {
-       int d0, d1, d2;
-
-       if (dest < src) {
-               if ((dest + n) < src)
-                        return memcpy(dest, src, n);
-               else
-                       __asm__ __volatile__(
-                               "rep\n\t"
-                               "movsb\n\t"
-                               : "=&c" (d0), "=&S" (d1), "=&D" (d2)
-                               :"0" (n),
-                                "1" (src),
-                                "2" (dest)
-                               :"memory");
-       } else {
-               if((src + n) < dest)
-                       return memcpy(dest, src, n);
-               else
-                       __asm__ __volatile__(
-                               "std\n\t"
-                               "rep\n\t"
-                               "movsb\n\t"
-                               "cld"
-                               : "=&c" (d0), "=&S" (d1), "=&D" (d2)
-                               :"0" (n),
-                                "1" (n-1+src),
-                                "2" (n-1+dest)
-                               :"memory");
-       }
-
-       return dest;
+       int d0,d1,d2,d3,d4,d5;
+       char *ret = dest;
+
+       __asm__ __volatile__(
+               /* Handle more 16bytes in loop */
+               "cmp $0x10, %0\n\t"
+               "jb     1f\n\t"
+
+               /* Decide forward/backward copy mode */
+               "cmp %2, %1\n\t"
+               "jb     2f\n\t"
+
+               /*
+                * movs instruction have many startup latency
+                * so we handle small size by general register.
+                */
+               "cmp  $680, %0\n\t"
+               "jb 3f\n\t"
+               /*
+                * movs instruction is only good for aligned case.
+                */
+               "mov %1, %3\n\t"
+               "xor %2, %3\n\t"
+               "and $0xff, %3\n\t"
+               "jz 4f\n\t"
+               "3:\n\t"
+               "sub $0x10, %0\n\t"
+
+               /*
+                * We gobble 16byts forward in each loop.
+                */
+               "3:\n\t"
+               "sub $0x10, %0\n\t"
+               "mov 0*4(%1), %3\n\t"
+               "mov 1*4(%1), %4\n\t"
+               "mov  %3, 0*4(%2)\n\t"
+               "mov  %4, 1*4(%2)\n\t"
+               "mov 2*4(%1), %3\n\t"
+               "mov 3*4(%1), %4\n\t"
+               "mov  %3, 2*4(%2)\n\t"
+               "mov  %4, 3*4(%2)\n\t"
+               "lea  0x10(%1), %1\n\t"
+               "lea  0x10(%2), %2\n\t"
+               "jae 3b\n\t"
+               "add $0x10, %0\n\t"
+               "jmp 1f\n\t"
+
+               /*
+                * Handle data forward by movs.
+                */
+               ".p2align 4\n\t"
+               "4:\n\t"
+               "mov -4(%1, %0), %3\n\t"
+               "lea -4(%2, %0), %4\n\t"
+               "shr $2, %0\n\t"
+               "rep movsl\n\t"
+               "mov %3, (%4)\n\t"
+               "jmp 11f\n\t"
+               /*
+                * Handle data backward by movs.
+                */
+               ".p2align 4\n\t"
+               "6:\n\t"
+               "mov (%1), %3\n\t"
+               "mov %2, %4\n\t"
+               "lea -4(%1, %0), %1\n\t"
+               "lea -4(%2, %0), %2\n\t"
+               "shr $2, %0\n\t"
+               "std\n\t"
+               "rep movsl\n\t"
+               "mov %3,(%4)\n\t"
+               "cld\n\t"
+               "jmp 11f\n\t"
+
+               /*
+                * Start to prepare for backward copy.
+                */
+               ".p2align 4\n\t"
+               "2:\n\t"
+               "cmp  $680, %0\n\t"
+               "jb 5f\n\t"
+               "mov %1, %3\n\t"
+               "xor %2, %3\n\t"
+               "and $0xff, %3\n\t"
+               "jz 6b\n\t"
+
+               /*
+                * Calculate copy position to tail.
+                */
+               "5:\n\t"
+               "add %0, %1\n\t"
+               "add %0, %2\n\t"
+               "sub $0x10, %0\n\t"
+
+               /*
+                * We gobble 16byts backward in each loop.
+                */
+               "7:\n\t"
+               "sub $0x10, %0\n\t"
+
+               "mov -1*4(%1), %3\n\t"
+               "mov -2*4(%1), %4\n\t"
+               "mov  %3, -1*4(%2)\n\t"
+               "mov  %4, -2*4(%2)\n\t"
+               "mov -3*4(%1), %3\n\t"
+               "mov -4*4(%1), %4\n\t"
+               "mov  %3, -3*4(%2)\n\t"
+               "mov  %4, -4*4(%2)\n\t"
+               "lea  -0x10(%1), %1\n\t"
+               "lea  -0x10(%2), %2\n\t"
+               "jae 7b\n\t"
+               /*
+                * Calculate copy position to head.
+                */
+               "add $0x10, %0\n\t"
+               "sub %0, %1\n\t"
+               "sub %0, %2\n\t"
+
+               /*
+                * Move data from 8 bytes to 15 bytes.
+                */
+               ".p2align 4\n\t"
+               "1:\n\t"
+               "cmp $8, %0\n\t"
+               "jb 8f\n\t"
+               "mov 0*4(%1), %3\n\t"
+               "mov 1*4(%1), %4\n\t"
+               "mov -2*4(%1, %0), %5\n\t"
+               "mov -1*4(%1, %0), %1\n\t"
+
+               "mov  %3, 0*4(%2)\n\t"
+               "mov  %4, 1*4(%2)\n\t"
+               "mov  %5, -2*4(%2, %0)\n\t"
+               "mov  %1, -1*4(%2, %0)\n\t"
+               "jmp 11f\n\t"
+
+               /*
+                * Move data from 4 bytes to 7 bytes.
+                */
+               ".p2align 4\n\t"
+               "8:\n\t"
+               "cmp $4, %0\n\t"
+               "jb 9f\n\t"
+               "mov 0*4(%1), %3\n\t"
+               "mov -1*4(%1, %0), %4\n\t"
+               "mov  %3, 0*4(%2)\n\t"
+               "mov  %4, -1*4(%2, %0)\n\t"
+               "jmp 11f\n\t"
+
+               /*
+                * Move data from 2 bytes to 3 bytes.
+                */
+               ".p2align 4\n\t"
+               "9:\n\t"
+               "cmp $2, %0\n\t"
+               "jb 10f\n\t"
+               "movw 0*2(%1), %%dx\n\t"
+               "movw -1*2(%1, %0), %%bx\n\t"
+               "movw %%dx, 0*2(%2)\n\t"
+               "movw %%bx, -1*2(%2, %0)\n\t"
+               "jmp 11f\n\t"
+
+               /*
+                * Move data for 1 byte.
+                */
+               ".p2align 4\n\t"
+               "10:\n\t"
+               "cmp $1, %0\n\t"
+               "jb 11f\n\t"
+               "movb (%1), %%cl\n\t"
+               "movb %%cl, (%2)\n\t"
+               ".p2align 4\n\t"
+               "11:"
+               : "=&c" (d0), "=&S" (d1), "=&D" (d2),
+                 "=r" (d3),"=r" (d4), "=r"(d5)
+               :"0" (n),
+                "1" (src),
+                "2" (dest)
+               :"memory");
+
+       return ret;
+
 }
 EXPORT_SYMBOL(memmove);
index ecacc4b..6d0f0ec 100644 (file)
 #undef memmove
 void *memmove(void *dest, const void *src, size_t count)
 {
-       unsigned long d0, d1, d2, d3;
-       if (dest < src) {
-               if ((dest + count) < src)
-                        return memcpy(dest, src, count);
-               else
-                       __asm__ __volatile__(
-                               "movq %0, %3\n\t"
-                               "shr $3, %0\n\t"
-                               "andq $7, %3\n\t"
-                               "rep\n\t"
-                               "movsq\n\t"
-                               "movq %3, %0\n\t"
-                               "rep\n\t"
-                               "movsb"
-                               : "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
-                               :"0" (count),
-                                "1" (src),
-                                "2" (dest)
-                               :"memory");
-       } else {
-               if((src + count) < dest)
-                       return memcpy(dest, src, count);
-               else
-                       __asm__ __volatile__(
-                               "movq %0, %3\n\t"
-                               "lea -8(%1, %0), %1\n\t"
-                               "lea -8(%2, %0), %2\n\t"
-                               "shr $3, %0\n\t"
-                               "andq $7, %3\n\t"
-                               "std\n\t"
-                               "rep\n\t"
-                               "movsq\n\t"
-                               "lea 7(%1), %1\n\t"
-                               "lea 7(%2), %2\n\t"
-                               "movq %3, %0\n\t"
-                               "rep\n\t"
-                               "movsb\n\t"
-                               "cld"
-                               : "=&c" (d0), "=&S" (d1), "=&D" (d2), "=r" (d3)
-                               :"0" (count),
-                                "1" (src),
-                                "2" (dest)
-                               :"memory");
-       }
-       return dest;
+       unsigned long d0,d1,d2,d3,d4,d5,d6,d7;
+       char *ret;
+
+       __asm__ __volatile__(
+               /* Handle more 32bytes in loop */
+               "mov %2, %3\n\t"
+               "cmp $0x20, %0\n\t"
+               "jb     1f\n\t"
+
+               /* Decide forward/backward copy mode */
+               "cmp %2, %1\n\t"
+               "jb     2f\n\t"
+
+               /*
+                * movsq instruction have many startup latency
+                * so we handle small size by general register.
+                */
+               "cmp  $680, %0\n\t"
+               "jb 3f\n\t"
+               /*
+                * movsq instruction is only good for aligned case.
+                */
+               "cmpb %%dil, %%sil\n\t"
+               "je 4f\n\t"
+               "3:\n\t"
+               "sub $0x20, %0\n\t"
+               /*
+                * We gobble 32byts forward in each loop.
+                */
+               "5:\n\t"
+               "sub $0x20, %0\n\t"
+               "movq 0*8(%1), %4\n\t"
+               "movq 1*8(%1), %5\n\t"
+               "movq 2*8(%1), %6\n\t"
+               "movq 3*8(%1), %7\n\t"
+               "leaq 4*8(%1), %1\n\t"
+
+               "movq %4, 0*8(%2)\n\t"
+               "movq %5, 1*8(%2)\n\t"
+               "movq %6, 2*8(%2)\n\t"
+               "movq %7, 3*8(%2)\n\t"
+               "leaq 4*8(%2), %2\n\t"
+               "jae 5b\n\t"
+               "addq $0x20, %0\n\t"
+               "jmp 1f\n\t"
+               /*
+                * Handle data forward by movsq.
+                */
+               ".p2align 4\n\t"
+               "4:\n\t"
+               "movq %0, %8\n\t"
+               "movq -8(%1, %0), %4\n\t"
+               "lea -8(%2, %0), %5\n\t"
+               "shrq $3, %8\n\t"
+               "rep movsq\n\t"
+               "movq %4, (%5)\n\t"
+               "jmp 13f\n\t"
+               /*
+                * Handle data backward by movsq.
+                */
+               ".p2align 4\n\t"
+               "7:\n\t"
+               "movq %0, %8\n\t"
+               "movq (%1), %4\n\t"
+               "movq %2, %5\n\t"
+               "leaq -8(%1, %0), %1\n\t"
+               "leaq -8(%2, %0), %2\n\t"
+               "shrq $3, %8\n\t"
+               "std\n\t"
+               "rep movsq\n\t"
+               "cld\n\t"
+               "movq %4, (%5)\n\t"
+               "jmp 13f\n\t"
+
+               /*
+                * Start to prepare for backward copy.
+                */
+               ".p2align 4\n\t"
+               "2:\n\t"
+               "cmp $680, %0\n\t"
+               "jb 6f \n\t"
+               "cmp %%dil, %%sil\n\t"
+               "je 7b \n\t"
+               "6:\n\t"
+               /*
+                * Calculate copy position to tail.
+                */
+               "addq %0, %1\n\t"
+               "addq %0, %2\n\t"
+               "subq $0x20, %0\n\t"
+               /*
+                * We gobble 32byts backward in each loop.
+                */
+               "8:\n\t"
+               "subq $0x20, %0\n\t"
+               "movq -1*8(%1), %4\n\t"
+               "movq -2*8(%1), %5\n\t"
+               "movq -3*8(%1), %6\n\t"
+               "movq -4*8(%1), %7\n\t"
+               "leaq -4*8(%1), %1\n\t"
+
+               "movq %4, -1*8(%2)\n\t"
+               "movq %5, -2*8(%2)\n\t"
+               "movq %6, -3*8(%2)\n\t"
+               "movq %7, -4*8(%2)\n\t"
+               "leaq -4*8(%2), %2\n\t"
+               "jae 8b\n\t"
+               /*
+                * Calculate copy position to head.
+                */
+               "addq $0x20, %0\n\t"
+               "subq %0, %1\n\t"
+               "subq %0, %2\n\t"
+               "1:\n\t"
+               "cmpq $16, %0\n\t"
+               "jb 9f\n\t"
+               /*
+                * Move data from 16 bytes to 31 bytes.
+                */
+               "movq 0*8(%1), %4\n\t"
+               "movq 1*8(%1), %5\n\t"
+               "movq -2*8(%1, %0), %6\n\t"
+               "movq -1*8(%1, %0), %7\n\t"
+               "movq %4, 0*8(%2)\n\t"
+               "movq %5, 1*8(%2)\n\t"
+               "movq %6, -2*8(%2, %0)\n\t"
+               "movq %7, -1*8(%2, %0)\n\t"
+               "jmp 13f\n\t"
+               ".p2align 4\n\t"
+               "9:\n\t"
+               "cmpq $8, %0\n\t"
+               "jb 10f\n\t"
+               /*
+                * Move data from 8 bytes to 15 bytes.
+                */
+               "movq 0*8(%1), %4\n\t"
+               "movq -1*8(%1, %0), %5\n\t"
+               "movq %4, 0*8(%2)\n\t"
+               "movq %5, -1*8(%2, %0)\n\t"
+               "jmp 13f\n\t"
+               "10:\n\t"
+               "cmpq $4, %0\n\t"
+               "jb 11f\n\t"
+               /*
+                * Move data from 4 bytes to 7 bytes.
+                */
+               "movl (%1), %4d\n\t"
+               "movl -4(%1, %0), %5d\n\t"
+               "movl %4d, (%2)\n\t"
+               "movl %5d, -4(%2, %0)\n\t"
+               "jmp 13f\n\t"
+               "11:\n\t"
+               "cmp $2, %0\n\t"
+               "jb 12f\n\t"
+               /*
+                * Move data from 2 bytes to 3 bytes.
+                */
+               "movw (%1), %4w\n\t"
+               "movw -2(%1, %0), %5w\n\t"
+               "movw %4w, (%2)\n\t"
+               "movw %5w, -2(%2, %0)\n\t"
+               "jmp 13f\n\t"
+               "12:\n\t"
+               "cmp $1, %0\n\t"
+               "jb 13f\n\t"
+               /*
+                * Move data for 1 byte.
+                */
+               "movb (%1), %4b\n\t"
+               "movb %4b, (%2)\n\t"
+               "13:\n\t"
+               : "=&d" (d0), "=&S" (d1), "=&D" (d2), "=&a" (ret) ,
+                 "=r"(d3), "=r"(d4), "=r"(d5), "=r"(d6), "=&c" (d7)
+               :"0" (count),
+                "1" (src),
+                "2" (dest)
+               :"memory");
+
+               return ret;
+
 }
 EXPORT_SYMBOL(memmove);