svn commit: trunk/uClibc: include libc/string libc/string/generic etc...

[Denys Vlasenko]

On Wednesday 22 July 2009 00:26, Bernhard Reutner-Fischer wrote:
> On Tue, Jul 21, 2009 at 03:36:03PM +0200, Denys Vlasenko wrote:
> >On Tue, Jul 21, 2009 at 4:01 AM, Mike Frysinger<vapier at gentoo.org> wrote:
> >>> Please try attached patch. For me it compiles. Resulting code
> >>> from memchr(buffer, '\n', pending):
> >>
> >> fixes building for me, but please unify the branches before committing
> >
> >Sorry. What does it mean "unify the branches"?
> 
> Only one insn changed between OPTIMIZE, i.e. please just:

Constraint changed too, from "i" (c) to "2" (c).
This was difficult to spot. This would be wrong
(it will probably build, but will contain extra
unneeded insn):

        void *edi;
        int ecx, eax;
        __asm__ __volatile__(
                "       jecxz   1f\n"
#if defined __OPTIMIZE__
                "       movb    %4, %%al\n" /* const c to %%al */
#endif
                "       repne; scasb\n"
                "       leal    -1(%%edi), %%edi\n"
                "       je      2f\n"
                "1:\n"
                "       xorl    %%edi, %%edi\n"
                "2:\n"
                : "=&D" (edi), "=&c" (ecx), "=&a" (eax)
                : "0" (s), "2" (c), "1" (count)
                /* : no clobbers */
        );
        return edi;


asm is error prone, and when I was coding the fix,
I decided that keeping -O0 and !-O0 implementations
totally separate is reducing that risk.

You do not need to mentally imagine how the code looks like
in both cases by parsing #ifdefs, and whether it is correct
- you already see both cases:

#if defined __OPTIMIZE__
        void *edi;
        int ecx, eax;
        __asm__ __volatile__(
                "       jecxz   1f\n"
                "       movb    %4, %%al\n" /* const c to %%al */
                "       repne; scasb\n"
                "       leal    -1(%%edi), %%edi\n"
                "       je      2f\n"
                "1:\n"
                "       xorl    %%edi, %%edi\n"
                "2:\n"
                : "=&D" (edi), "=&c" (ecx), "=&a" (eax)
                : "0" (s), "i" (c), "1" (count)
                /* : no clobbers */
        );
        return edi;
#else
        /* With -O0, gcc can't figure out how to encode CONST c
         * as an immediate operand. Generating slightly bigger code
         * (usually "movl CONST,%eax", 3 bytes bigger than needed):
         */
        void *edi;
        int ecx, eax;
        __asm__ __volatile__(
                "       jecxz   1f\n"
                "       repne; scasb\n"
                "       leal    -1(%%edi), %%edi\n"
                "       je      2f\n"
                "1:\n"
                "       xorl    %%edi, %%edi\n"
                "2:\n"
                : "=&D" (edi), "=&c" (ecx), "=&a" (eax)
                : "0" (s), "2" (c), "1" (count)
                /* : no clobbers */
        );
        return edi;
#endif


But it's wordy, yes.
--
vda