Incorrect behavior on C function that returns struct which contains array of one or two doubles

[tatjam]

Zig Version

0.14.0-dev.1349+6a21875dd

Steps to Reproduce and Observed Behavior

Consider the following C struct and function:

// test.h
typedef struct {
  double dat[2];
} complex_number;

complex_number test_function();

// test.c
complex_number test_function() {
  complex_number out;

  out.dat[0] = 3;
  out.dat[1] = -6;

  return out;
}

Equivalently, the test_function can be implemented in Zig as follows:

// main.zig
const c = @cImport(@cInclude("test.h"));

pub fn zig_equiv() c.complex_number {
	var out: c.complex_number = undefined;
	out.dat[0] = 3;
	out.dat[1] = -6;
	return out;
}

In either case, the functions may be invoked as follows:

// main.zig
const std = @import("std");

pub fn main() !void {
  const res = c.test_function();
  //const res = zig_equiv();

  std.log.info("{}, {}", .{res.dat[0], res.dat[1]});
}

The code is built on an Ubuntu 22.04 without additional flags via zig build. The behavior is as follows:

• If the C code is called, the log will return nonsense values (that vary each run).
• If the Zig code is called, the expected values are printed
• If the struct's array is adjusted to contain more values (those are simply left uninitialized), the code starts working correctly on either C or Zig implementation.
• If compiling with ReleaseFast, incorrect behavior still happens, but in this case, instead of random values, 0 is printed.

I attach a minimum reproducible case (minimum_case.zip) that contains the code I've used. Should be as straightforward as downloading the zip and running zig build run.

For context, the error has come up while trying to wrap a C library that uses complex values, stored as this struct of an array of two elements.

Expected Behavior

For the struct to be correctly returned.

[rohlem]

Sounds like a C ABI issue, which is based on the CPU architecture of your system (probably either x86_64 or aarch64?), but also specific CPU features.
You could also try adding -mcpu baseline as a compile flag and see whether the issue persists.

For x86_64 there used to be an old issue #1481 , which has since been closed.
For aarch64 there is #14908 , although it seems primarily Windows is affected.
Either way, it working in Zig 0.13.0 means this is a regression, which should be tagged (and I believe may be prioritized before the next release, 0.14.0).
(Should also be tagged miscompilation fwict.)

[tatjam]

I'm indeed running x86_64, forgot to mention. The case appears to fall within "x86_64: struct & union return values <= 16 bytes"

[tatjam]

Speaking of which, I checked again and this case is NOT working on Zig 0.13, I accidentally ran the code with the struct containing more than 2 doubles. Going to update the main post.

[vesim987]

Just from quick glance, on x86_64:
zig:

declare { i64, i64 } @test_function() #5

c:

define dso_local { double, double } @test_function() #0 {  

this is causing zig to read the return value from rdx and rax instead of xmm0 and xmm1.

zig/src/arch/x86_64/abi.zig

Lines 308 to 319 in 2111f4c

	.array => {
	const ty_size = ty.abiSize(zcu);
	if (ty_size <= 8) {
	result[0] = .integer;
	return result;
	}
	if (ty_size <= 16) {
	result[0] = .integer;
	result[1] = .integer;
	return result;
	}
	return memory_class;

I think this piece of code need special handling of floats