perf_counter.h

/*
 * ====== About Performance Counters ======
 *
 * Technically a performance counter is an ordinary C/C++ variable.
 * It needs a declaration
 *
 *   DECLARE_PERF_COUNTER(int, my_counter, "my_counter"),
 *
 * ...and the definition:
 *
 *   DEFINE_PERF_COUNTER(int, my_counter, "my_counter").
 *
 *
 * The above statements declare/define 'int my_counter' variable.  The
 * code accesses my_counter variable like any other variable.
 * Additionally the performance counter directory is build automatically
 * enabling runtime enumeration and lookups by a string name.
 *
 * Last but not least counters are grouped in a compact page-size
 * aligned memory range; it is possible to map a shared memory segment
 * on top of the counters enabling low-overhead performance data
 * collection with an external tool.
 *
 */
#ifndef PERF_COUNTER_H__
#define PERF_COUNTER_H__


#define DECLARE_PERF_COUNTER(type, name, str_name) \
    extern type name __attribute__((__visibility__("hidden")));


#define DEFINE_PERF_COUNTER(type, name, str_name) \
    DEFINE_HIDDEN_VARIABLE_IN_SECTION__(type,name,".PERFCOUNTER,\"aw\",@nobits#") \
    static const perf_counter_desc_t name ## _descriptor__ = { \
        str_name, (perf_counter_t *) &name, 0 }; \
    static const perf_counter_desc_t *name ## _descriptor_ref__ \
    __attribute__((__used__,__section__(".PERFCOUNTERDESC"))) = \
        &name ## _descriptor__;


/* Implementation follows (Here Be Dragons) */


#pragma GCC visibility push(hidden)
#ifdef __cplusplus__
extern "C" {
#endif /* __cplusplus__ */


typedef struct perf_counter_s          perf_counter_t;
typedef struct perf_counter_desc_s     perf_counter_desc_t;


/* Provided by ld, see linkerscript */
extern perf_counter_t             perf_counter_begin__;
extern perf_counter_t             perf_counter_end__;
extern const perf_counter_desc_t *perf_counter_desc_begin__;
extern const perf_counter_desc_t *perf_counter_desc_end__;


struct perf_counter_s
{
    int            reserved__;
};


struct perf_counter_desc_s
{
    const char    *name;
    perf_counter_t*ptr;
    int            flags;
};


/*
 * DEFINE_HIDDEN_VARIABLE_IN_SECTION__(type,name,section) - a variable
 * definition with a section attribute in GCC, ungodly mess in the clang
 * variant.
 *
 * The reason for this insanity?  We want to place performance counters
 * in the very special section, a @nobits one (@nobits sections go to
 * .bss).  GCC forces a wrong section type (@progbits), however a
 * workaround exists since GCC fails to escape section names properly.
 *
 * Clang breaks the workaround  (http://llvm.org/bugs/show_bug.cgi?id=16232)
 * the only viable option was to generate the correct definition with
 * inline assembly.
 */
#ifdef __clang__
/*
 * Bonus insanity: it was necessary to wrap assembly statement in a
 * dummy function (don't worry the dummy doesn't make it through into
 * the final executable thanks to the magic of ".discard" section)
 * (http://psomas.wordpress.com/2011/02/25/inline-assembly-gcc-attributes-linker-scripts-and-a-kernel-macro/)
 *
 * Even more jumping through hoops due to the need for unique function
 * names (hello, __COUNTER__).
 */
#define DEFINE_HIDDEN_VARIABLE_IN_SECTION__(type,name,section) \
    DEFINE_HIDDEN_VARIABLE_IN_SECTION_INTERNAL__(type,name,section,__COUNTER__)
#define DEFINE_HIDDEN_VARIABLE_IN_SECTION_INTERNAL__(type,name,section,i) \
    extern type name __attribute__((__visibility__("hidden"))); \
    static void __attribute__((__used__,__section__(".discard"))) \
    CONCAT__(declare_variable_in_section__, i) () { \
        __asm__( \
        ".type\t" #name ", @object;\n" \
        "\t.pushsection\t" section "\n" \
        "\t.hidden\t" #name "\n" \
        "\t.globl\t" #name "\n" \
        "\t.align\t%c0;\n" \
        #name ":\n" \
        "\t.zero\t%c1\n" \
        "\t.size\t" #name ", %c1\n" \
        "\t.popsection" \
        ::"i"(__alignof__(type)), "i"(sizeof(type))); \
    }
#define CONCAT__(a,b) a ## b
#elif __GNUC__
/* The real cakewalk! */
#define DEFINE_HIDDEN_VARIABLE_IN_SECTION__(type,name,section) \
    type name \
    __attribute__((__section__(section),__visibility__("hidden")));
#else
#error Weird compiler
#endif


#ifdef __cplusplus__
} /* extern "C" { */
#endif /* __cplusplus__ */
#pragma GCC visibility pop

#endif /* PERF_COUNTER_H__ */