Description
Zig Version
0.15.0-dev.847+850655f06
Steps to Reproduce and Observed Behavior
I run zig's tests locally with -fqemu and I've noticed some unusual and persistent failures on the s390x-linux target. I see failures in hypot, gamma and complex tests. I've only narrowed down the hypot one at this point, I can dig into the others if that's useful. From what I can see its just f32 floats that are failing the hypot test case, and its the "easy" test (not a boundary test case). The f16, f64, f80, and f128 floats pass. And the test passes on x86*, arm and mips at least.
// Isolated from "hypot.precise" test case in lib/std/math/hypot.zig
//
// $ zig run main.zig
// $ zig run --test-cmd qemu-s390x --test-cmd-bin -target s390x-linux main.zig
const std = @import("std");
pub fn main() !void {
checkhypot(f16);
checkhypot(f32);
checkhypot(f64);
checkhypot(f80);
checkhypot(f128);
}
fn checkhypot(T: type) void {
const a: T = 0.0;
const b: T = -1.2;
const expected: T = 1.2;
const computed = @sqrt(a * a + b * b);
// This is pulled from hypot.zig and should be how the 'f32' case is handled (But its fine here?)
const computed64 = if (T == f80 or T == f128) 0 else @sqrt(@mulAdd(f64, a, a, @as(f64, b) * b));
const actual = std.math.hypot(a, b);
const tolerance = @sqrt(std.math.floatEps(T));
if (! std.math.approxEqRel(T, expected, actual, tolerance)) {
std.debug.print("FAIL: {any} a={} b={} actual={} expected={} computed={} computed64={}\n", .{
T, a, b, actual, expected, computed, computed64
});
} else {
std.debug.print("PASS: {any}\n", .{ T });
}
}
If I run this locally (x86_64-linux) it passes:
$ zig run main.zig
PASS: f16
PASS: f32
PASS: f64
PASS: f80
PASS: f128
The s390x target just fails the f32 version:
$ zig run --test-cmd qemu-s390x --test-cmd-bin -target s390x-linux main.zig
PASS: f16
FAIL: f32 a=0e0 b=-1.2e0 actual=1.732051e-1 expected=1.2e0 computed=1.2e0 computed64=1.2000000476837158e0
PASS: f64
PASS: f80
PASS: f128
The "actual" value is very wrong compared to "expected", so its not just a floating-point tolerance issue.
This could be a qemu bug, but that seems unlikely at this point. If there are any more useful details I can squeeze out of Zig, let me know.
Expected Behavior
PASS: f32 on s390x-linux target.