ww/test/wcc/989_ffivariadic_run.c

/*
 * 989_ffivariadic_run — C1 (catB-54): a ww caller of a C variadic function
 * (`@symbol("f") fn f(a: i64, ...) f64;`) must set the SysV AL register to
 * the number of XMM regs used to pass the variadic FLOAT args. The C callee
 * gates its xmm-save-area stores on `test %al,%al`, so a wrong AL makes
 * va_arg(double) read garbage.
 *
 * THE BUG (cat-A silent miscompile, byte-id-blind): cgen.c hardcoded AL=0
 * (`XORQ AX,AX`) at the variadic-call site — correct only for a zero-float
 * variadic call. THE FIX: emit AL = the XMM cursor `fi` (the count of float
 * args placed in XMM regs). Ref SysV §3.5.7, ref/qbe/amd64/sysv.c:384.
 *
 * RUNTIME gate (byte-id can never see AL correctness): each row builds a ww
 * caller that calls the C fixture `double fixture(long n, ...)` (a
 * va_arg(double) summer, test/wcc/data/ffivariadic/fixture.c, linked from
 * libffifix.a) and asserts the returned sum. Runs on BOTH the cstage `ww`
 * and wwstage `ww_ww` drivers (C2): AL is byte-id-blind, so a wwstage fi /
 * f32-promotion divergence is caught only by a wrong runtime sum here.
 *
 * NON-VACUITY DEVIATION (reported to lead): the spec's `fixture(2,1.0,2.0)`
 * is VACUOUS on this box — with AL=0 the two skipped xmm slots happen to
 * alias stale stack that already holds 1.0/2.0, so the 2-float call returns
 * the correct 3.0 even unfixed. At 3+ floats the coincidence breaks: AL=0
 * deterministically returns the wrong sum. Every row below uses >=3 floats,
 * so reverting the fix to `XORQ AX,AX` FAILS this test (proven). A 2-float
 * row would pass both ways and prove nothing.
 */
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <sys/stat.h>
#include <sys/wait.h>

static int
runwait(const char *cmd)
{
	int rc = system(cmd);
	if (rc == -1) return -1;
	if (WIFEXITED(rc)) return WEXITSTATUS(rc);
	return -1;
}

struct row {
	const char *label;
	const char *src;
	int want_exit;	/* 0 == sum matched */
};

static const struct row rows[] = {
	/* 3 floats: 1+2+3 == 6. The smallest non-vacuous count (see header). */
	{ "three",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let r: f64 = fixture(3, 1.0, 2.0, 3.0);\n"
	  "	if (r == 6.0) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },

	/* 5 floats, fewer than the 8 XMM arg regs: 1+2+3+4+5 == 15. */
	{ "five",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let r: f64 = fixture(5, 1.0, 2.0, 3.0, 4.0, 5.0);\n"
	  "	if (r == 15.0) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },

	/* 8 floats == all XMM arg regs (AL caps at 8): 1+..+8 == 36. */
	{ "eight",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let r: f64 = fixture(8, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0);\n"
	  "	if (r == 36.0) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },

	/* 3 f32 args (#14): C default arg promotion widens each to f64, so
	 * the callee's va_arg(double) reads 1.5+2.5+3.0 == 7.0. Values are
	 * exact in f32, so the f64 compare is exact. Unpromoted (MOVSS, 4B)
	 * the high 4B of each 8B slot is stale stack, so va_arg(double)
	 * pulls garbage and the sum misses 7.0 — this row FAILS pre-fix. */
	{ "f32three",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let r: f64 = fixture(3, 1.5: f32, 2.5: f32, 3.0: f32);\n"
	  "	if (r == 7.0) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },

	/* Sharper than f32three: 0.1 is INEXACT in f32, so the f32-rounded
	 * value differs from the f64 literal 0.1. The expected side `(x: f64)`
	 * is the SAME f32 var widened at runtime (CVTSS2SD), so equality holds
	 * iff the variadic arg carried the f32-rounded value promoted to f64 —
	 * NOT the original f64 literal and NOT stale-high-bit garbage. Both a
	 * no-promote (MOVSS) and a hypothetical direct-f64 pass would miss it.
	 * Single f32 arg also exercises the fi==1 boundary. */
	{ "f32inexact",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let x: f32 = 0.1: f32;\n"
	  "	let r: f64 = fixture(1, x);\n"
	  "	if (r == (x: f64)) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },

	/* Mixed f32 + f64 in one variadic call: only the f32 args (1.5, 3.5)
	 * promote; the f64 arg (2.0) passes at its native width and must NOT be
	 * double-promoted or skewed. 1.5+2.0+3.5 == 7.0, exact in both widths. */
	{ "f32mixed",
	  "package main;\n"
	  "@symbol(\"fixture\") fn fixture(n: i64, ...) f64;\n"
	  "export fn main() int = {\n"
	  "	let r: f64 = fixture(3, 1.5: f32, 2.0, 3.5: f32);\n"
	  "	if (r == 7.0) { return 0; };\n"
	  "	return 1;\n"
	  "};\n",
	  0 },
};

/* run_build — build+run `src` via cstage `driver`, linking libffifix.a from
 * `libdir`. Returns the binary's exit code, or -1 on a build failure. */
static int
run_build(const char *driver, const char *libdir, const struct row *r, int i)
{
	char src[64], tmpdir[64], cmd[1024];
	snprintf(src, sizeof src, "/tmp/ffivar_%d_%d.ww", getpid(), i);
	snprintf(tmpdir, sizeof tmpdir, "/tmp/ffivar_%d_d_%d", getpid(), i);

	FILE *f = fopen(src, "wb");
	if (!f) return -2;
	fputs(r->src, f);
	fclose(f);

	mkdir(tmpdir, 0755);
	snprintf(cmd, sizeof cmd,
	    "cd %s && %s build -L%s -lffifix %s 2>/dev/null",
	    tmpdir, driver, libdir, src);
	int brc = runwait(cmd);

	const char *base = strrchr(src, '/');
	base = base ? base + 1 : src;
	char outbin[128];
	snprintf(outbin, sizeof outbin, "%s/%s", tmpdir, base);
	char *dot = strrchr(outbin, '.');
	if (dot && strcmp(dot, ".ww") == 0) *dot = '\0';

	int got = -1;
	if (brc == 0) got = runwait(outbin);

	unlink(src); unlink(outbin); rmdir(tmpdir);
	return brc == 0 ? got : -1;
}

int
main(void)
{
	const char *bin = getenv("BIN");
	if (!bin) bin = "out/bin";
	char absbin[1024];
	if (bin[0] != '/') {
		char cwd[1024];
		if (getcwd(cwd, sizeof cwd) == NULL) return 1;
		snprintf(absbin, sizeof absbin, "%s/%s", cwd, bin);
		bin = absbin;
	}

	char cdrv[1024], wdrv[1024], libdir[1024];
	snprintf(cdrv, sizeof cdrv, "%s/ww", bin);
	snprintf(wdrv, sizeof wdrv, "%s/ww_ww", bin);
	/* libffifix.a lives beside $(BIN) under $(OUT)/ffivariadic — the
	 * Makefile builds it there as a prereq of this test binary. */
	snprintf(libdir, sizeof libdir, "%s/../ffivariadic", bin);

	/* C2: run each row on BOTH the cstage `ww` and the wwstage `ww_ww`
	 * driver. AL correctness is byte-id-blind, so a wwstage fi/promotion
	 * divergence is invisible to the 990-997 gates but caught here as a
	 * wrong sum (nonzero exit). wwstage is access-gated like the other
	 * dual-stage runtime tests (989_chainidx_run) so a cstage-only tree
	 * still runs the cstage rows. */
	struct { const char *name; const char *drv; int gated; }
	drivers[] = {
		{ "cstage",  cdrv, 0 },
		{ "wwstage", wdrv, 1 },
		{ NULL, NULL, 0 },
	};

	int n = (int)(sizeof rows / sizeof rows[0]);
	int total = 0, fail = 0;

	for (int d = 0; drivers[d].name; d++) {
		if (drivers[d].gated && access(drivers[d].drv, X_OK) != 0) {
			fprintf(stderr, "ffivariadic: skip %s (no %s)\n",
			    drivers[d].name, drivers[d].drv);
			continue;
		}
		for (int i = 0; i < n; i++) {
			total++;
			int got = run_build(drivers[d].drv, libdir, &rows[i], i);
			if (got != rows[i].want_exit) {
				fprintf(stderr,
				    "ffivariadic[%s][%s]: exit=%d want=%d\n",
				    drivers[d].name, rows[i].label, got,
				    rows[i].want_exit);
				fail++;
			}
		}
	}

	if (fail) {
		fprintf(stderr, "ffivariadic: %d/%d fixtures failed\n",
		    fail, total);
		return 1;
	}
	printf("ffivariadic: %d/%d ok\n", total, total);
	return 0;
}