Validate SHRN/FCMP check for all-zero boolean bitmap works in the presence of FZ/FIZ/AH

ftztest.c

/* for FreeBSD, some patching required to run on Linux */
#include <arm_neon.h>
#include <machine/armreg.h>
#include <stdbool.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

static void
usage(char *argv0)
{
	fprintf(stderr, "usage: %s [-afi]\n  -a for AH mode\n  -f for FZ mode\n  -i for FIZ mode\n", argv0);
	exit(EXIT_FAILURE);
}

static int
veq_zero(uint8x16_t v)
{
#ifdef __arm__
	union { uint8x8_t v; double d; } narrowed;

	/* narrow each byte to a nibble */
	narrowed.v = vshrnnarrowed.v = vshrn_n_u16(vreinterpretq_u16_u8(v), 4);

	return (narrowed.d == 0.0);
#else /* AArch64 */
	uint8x8_t narrowed;

	/* narrow each byte to a nibble */
	narrowed = vshrn_n_u16(vreinterpretq_u16_u8(v), 4);

	/* check if that vector is all zero */
	return (vdupd_lane_f64(vreinterpret_f64_u16(narrowed), 0) == 0.0);
#endif
}

/* check if veq_zero works correctly for all possible combinations of input */
static void
run_test(void)
{
	size_t i, j;
	int out;
	unsigned char v[16];

	for (i = 0; i <= 0xffff; i++) {
		for (j = 0; j < 16; j++)
			v[j] = i & 1 << j ? 0xff : 0x00;

		out = veq_zero(vld1q_u8(v));
		if (veq_zero(vld1q_u8(v)) != (i == 0))
			printf("%04zx: wanted %d got %d\n", i, i == 0, out);
	}
}

int main(int argc, char *argv[])
{
	uint64_t fpcr;
	int ch;
	bool ah = false, fz = false, fiz = false;

	while (ch = getopt(argc, argv, "afi"), ch != -1)
		switch (ch) {
		case 'a':
			ah = true;
			break;

		case 'f':
			fz = true;
			break;

		case 'i':
			fiz = true;
			break;

		default:
			usage(argv[0]);
		}

	fpcr = READ_SPECIALREG(FPCR);
	fpcr &= ~0x01000003ULL;

	if (fiz)
		fpcr |= 1;

	if (ah)
		fpcr |= 2;

	if (fz)
		fpcr |= 0x01000000;

	WRITE_SPECIALREG(FPCR, fpcr);

	run_test();
}