Compiler bug

[james-elliot]

Zig Version

0.11.0-dev.815+88b49ed00

Steps to Reproduce and Observed Behavior

In the following code, the program compiles and runs properly when NB_BITS is 27 or less, but the compilation fails if it is 28 with a relocation problem:

zig build-exe p4.zig
error: ld.lld: p4.o:(function start.posixCallMainAndExit: .text+0x64d): relocation R_X86_64_32S out of range: 2149950032 is not in [-2147483648, 2147483647]

const NB_BITS: u8 = 28;

const std = @import("std");
const stdout = std.io.getStdOut().writer();

const HASH_MASK: u64 = (1 << NB_BITS) - 1;
const HASH_SIZE: usize = 1 << NB_BITS;
var hashes: [HASH_SIZE]u64 = undefined;

fn retrieve(hv: u64) bool {
    const ind: usize = (hv & HASH_MASK);
    if (hashes[ind] == hv) return true else return false;
}

pub fn main() !void {
    for (hashes) |*a| a.* = 0;
    if (retrieve(0)) try stdout.print("{d}\n", .{0});
}

Expected Behavior

Correct compilation...

[FanShupei]

I'd like to say this is not zig's bug, this is the intended behavior. The related concept is [code model][https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html]. In gcc, it's controlled by -mcmodel=small|medium|large. The default model is small model, all code and data must live in 2GB. When NB_BITS is set to 28 or higher, hashes will exceed 2GB. That's the problem.

However, currently zig's support for code model is very limited, only supporting default and tiny. If zig adds support for medium and large code model, the code should run in proper code model.

[Vexu]

I'd like to say this is not zig's bug, this is the intended behavior. The related concept is [code model][https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html]. In gcc, it's controlled by -mcmodel=small|medium|large. The default model is small model, all code and data must live in 2GB. When NB_BITS is set to 28 or higher, hashes will exceed 2GB. That's the problem.

It's still a bug in that there should be a compiler error before it reaches the linker.

However, currently zig's support for code model is very limited, only supporting default and tiny. If zig adds support for medium and large code model, the code should run in proper code model.

Medium and large are supported in addition to small and kernel.

  -mcmodel=[default|tiny|   Limit range of code and data virtual addresses
            small|kernel|
            medium|large]

[FanShupei]

It's still a bug in that there should be a compiler error before it reaches the linker.

I agree with you. gcc and clang do nothing and just defer the error to linking. If zig could detect this then give a nicer message instead of a cryptic linking error, this is definitely an improvement. But I'm afraid that it's hard to solve perfectly. Thinking an extreme case, two object files each contains a 1GB array, it will only explode when they are linking together.

Medium and large are supported in addition to small and kernel.

You are right. I was just using zig build-exe -h | grep mcmodel, so I didn't see the last two lines, Sorry.

[Vexu]

I agree with you. gcc and clang do nothing and just defer the error to linking. If zig could detect this then give a nicer message instead of a cryptic linking error, this is definitely an improvement. But I'm afraid that it's hard to solve perfectly. Thinking an extreme case, two object files each contains a 1GB array, it will only explode when they are linking together.

Sure but a lot of the times a compilation will only contain Zig code which is all in one compilation unit in which this should be possible to detect.