Re: [PATCH v1] lib/vsprintf.c: Even faster decimal conversion

[Jeff Epler <jepler@unpythonic.net>]

Since you asked about big-endian systems I also built your test program
for the armeb architecture -- which involved hacking up the test harness
fairly heavily to not require libc -- and ran the result in qemu.

Actually, it hasn't finished after 2 hours of (qemu) CPU time, but I can
tell from the output that the first phase has completed successfully but
the second hasn't finished (because it's printed 1 and 2 but not yet 3).
I am pretty sure this is because it's just rather slow in emulation, and
obviously performance figures are going to be useless.

Anyway, I built this on my armhf machine with
    gcc-4.8 -O -o testeb linux32.c rv32.c verify.c -nostdlib -mbig-endian
copied it to my fastest x86_64 desktop and ran it with
    qemu-armeb ./a.out || echo fail

#include <stdio.h>
#include <stdlib.h>
#include <limits.h>
#include <pthread.h>
#include <string.h>

#include "common.h"

#define NTHR 1

#define LO_START 10ULL
#define LO_STOP  10000000000ULL
#define HI_START ULLONG_MAX
#define HI_STOP  ULLONG_MAX - LO_STOP

static unsigned long long lenfreq[NTHR][32];

#include <unistd.h>
#include <sys/syscall.h>
__attribute__((noinline))
static void do_exit(int i) {
    register int i0 asm("r0") = i;
    register int ss asm("r7") = SYS_exit;
    __asm__ __volatile__("swi 0x0" : : "r"(ss), "r"(i0));
}

__attribute__((noinline))
static void do_write(int fd, const char *buf, size_t n) {
    register int i0 asm("r0") = fd;
    register const char *i1 asm("r1") = buf;
    register size_t i2 asm("r2") = n;
    register int ss asm("r7") = SYS_write;
    __asm__ __volatile__("swi 0x0" : : "r"(ss), "r"(i0), "r"(i1), "r"(i2));
}

int memcmp(const void * va, const void * vb, size_t sz) {
    unsigned char *a, *b, aa, bb;
    while(sz--) {
        int diff = *a++ - *b++;
        if(diff) return diff;
    }
    return 0;
}

static int do_check(unsigned long long n, unsigned idx)
{
	char buf1[24];
	char buf2[24];
	int len1, len2;

	len1 = linux_put_dec(buf1, n) - buf1;
	len2 = rv_put_dec(buf2, n) - buf2;
	if (len1 != len2 || memcmp(buf1, buf2, len1)) {
		return -1;
	}
	lenfreq[idx][len1]++;
	return 0;
}

static void *check(void *arg)
{
	unsigned long idx = (unsigned long)arg;
	unsigned long long n;

        do_write(1, "1\n", 2);
	for (n = LO_START; n % NTHR != idx; ++n)
		;
	for (; n <= LO_STOP; n += NTHR) {
		if (do_check(n, idx))
			return (void*) -1;
	}

        do_write(1, "2\n", 2);
	for (n = HI_START; n % NTHR != idx; --n)
		;
	for (; n >= HI_STOP; n -= NTHR) {
		if (do_check(n, idx))
			return (void*) -1;
	}

	/*
	 * This will also visit a few one-digit numbers, but both the
	 * old and new code actually handle that just fine for
	 * non-zero n (it's just irrelevant because all callers of
	 * put_dec take a shortcut for n < 10).
	 */
        do_write(1, "3\n", 2);
	n = 2*idx + 1;
	do {
		if (do_check(n, idx))
			return (void*) -1;
		n *= 17179869185ull;
	} while (n != 2*idx + 1);

	return NULL;
}

int _start(void)
{
    do_write(1, ".\n", 2);
    do_exit(!!check(0));
}