URL https://opencores.org/ocsvn/or1k/or1k/trunk

Subversion Repositories or1k

[/] [or1k/] [trunk/] [rc203soc/] [sw/] [uClinux/] [arch/] [sparc/] [lib/] [strlen.S] - Blame information for rev 1765

Details | Compare with Previous | View Log


/* strlen.S: Sparc optimized strlen().
 *
 * This was hand optimized by davem@caip.rutgers.edu from
 * the C-code in GNU-libc.
 */
 
#include 
 
#define LO_MAGIC 0x01010101
#define HI_MAGIC 0x80808080
 
        .align 4
        .global C_LABEL(strlen)
C_LABEL(strlen):
        mov     %o0,%o1
        andcc   %o0,3,%g0               ! and with %o0 so no dependency problems
        be      scan_words
         sethi  %hi(HI_MAGIC),%g2       ! common case and most Sparcs predict taken
 
        ldsb    [%o0],%g2
still_not_word_aligned:
        cmp     %g2,0
        bne,a   1f
         add    %o0,1,%o0
 
        /* Ok, so there are tons of quick interlocks above for the
         * < 4 length string unaligned... not too common so I'm not
         * very concerned.
         */
        retl
         sub    %o0,%o1,%o0
 
1:
        andcc   %o0,3,%g0
        bne,a   still_not_word_aligned
         ldsb   [%o0],%g2
 
        /* HyperSparc executes each sethi/or pair in 1 cycle. */
        sethi   %hi(HI_MAGIC),%g2
scan_words:
        or      %g2,%lo(HI_MAGIC),%o3
        sethi   %hi(LO_MAGIC),%g3
        or      %g3,%lo(LO_MAGIC),%o2
next_word:
        ld      [%o0],%g2               ! no dependencies
next_word_preloaded:
        sub     %g2,%o2,%g2             ! lots of locks here
        andcc   %g2,%o3,%g0             ! and I dont like it...
        be      next_word
         add    %o0,4,%o0
 
        /* Check every byte. */
byte_zero:
        ldsb    [%o0-4],%g2
        cmp     %g2,0
        bne     byte_one
         add    %o0,-4,%g3
 
        retl
         sub    %g3,%o1,%o0
 
byte_one:
        ldsb    [%o0-3],%g2
        cmp     %g2,0
        bne,a   byte_two_and_three
         ldsb   [%o0-2],%g2
 
        sub     %g3,%o1,%o0
        retl
         add    %o0,1,%o0
 
byte_two_and_three:
        cmp     %g2,0
        be,a    found_it
         sub    %g3,%o1,%o0
 
        ldsb    [%o0-1],%g2
        cmp     %g2,0
        bne,a   next_word_preloaded
         ld     [%o0],%g2
 
        sub     %g3,%o1,%o0
        retl
         add    %o0,3,%o0
 
found_it:
        retl
         add    %o0,2,%o0

Line No.	Rev	Author	Line
1	1624	jcastillo	`/* strlen.S: Sparc optimized strlen().`
2			`*`
3			`* This was hand optimized by davem@caip.rutgers.edu from`
4			`* the C-code in GNU-libc.`
5			`*/`
6
7			`#include`
8
9			`#define LO_MAGIC 0x01010101`
10			`#define HI_MAGIC 0x80808080`
11
12			`.align 4`
13			`.global C_LABEL(strlen)`
14			`C_LABEL(strlen):`
15			`mov %o0,%o1`
16			`andcc %o0,3,%g0 ! and with %o0 so no dependency problems`
17			`be scan_words`
18			`sethi %hi(HI_MAGIC),%g2 ! common case and most Sparcs predict taken`
19
20			`ldsb [%o0],%g2`
21			`still_not_word_aligned:`
22			`cmp %g2,0`
23			`bne,a 1f`
24			`add %o0,1,%o0`
25
26			`/* Ok, so there are tons of quick interlocks above for the`
27			`* < 4 length string unaligned... not too common so I'm not`
28			`* very concerned.`
29			`*/`
30			`retl`
31			`sub %o0,%o1,%o0`
32
33			`1:`
34			`andcc %o0,3,%g0`
35			`bne,a still_not_word_aligned`
36			`ldsb [%o0],%g2`
37
38			`/* HyperSparc executes each sethi/or pair in 1 cycle. */`
39			`sethi %hi(HI_MAGIC),%g2`
40			`scan_words:`
41			`or %g2,%lo(HI_MAGIC),%o3`
42			`sethi %hi(LO_MAGIC),%g3`
43			`or %g3,%lo(LO_MAGIC),%o2`
44			`next_word:`
45			`ld [%o0],%g2 ! no dependencies`
46			`next_word_preloaded:`
47			`sub %g2,%o2,%g2 ! lots of locks here`
48			`andcc %g2,%o3,%g0 ! and I dont like it...`
49			`be next_word`
50			`add %o0,4,%o0`
51
52			`/* Check every byte. */`
53			`byte_zero:`
54			`ldsb [%o0-4],%g2`
55			`cmp %g2,0`
56			`bne byte_one`
57			`add %o0,-4,%g3`
58
59			`retl`
60			`sub %g3,%o1,%o0`
61
62			`byte_one:`
63			`ldsb [%o0-3],%g2`
64			`cmp %g2,0`
65			`bne,a byte_two_and_three`
66			`ldsb [%o0-2],%g2`
67
68			`sub %g3,%o1,%o0`
69			`retl`
70			`add %o0,1,%o0`
71
72			`byte_two_and_three:`
73			`cmp %g2,0`
74			`be,a found_it`
75			`sub %g3,%o1,%o0`
76
77			`ldsb [%o0-1],%g2`
78			`cmp %g2,0`
79			`bne,a next_word_preloaded`
80			`ld [%o0],%g2`
81
82			`sub %g3,%o1,%o0`
83			`retl`
84			`add %o0,3,%o0`
85
86			`found_it:`
87			`retl`
88			`add %o0,2,%o0`

Browse

Tools

Subversion Repositories or1k

[/] [or1k/] [trunk/] [rc203soc/] [sw/] [uClinux/] [arch/] [sparc/] [lib/] [strlen.S] - Blame information for rev 1765