Fix load alignment being applied to referenced value

[sapir]

In some code I compiled, the align argument (edit: in Builder::load) wasn't the same as the pointer's regular alignment, and gcc ended up dereferencing the wrong address. So I think the align argument is meant to only apply to the loaded value and not to the one that's being read.

For comparison, rustc_codegen_llvm does the following:

let load = llvm::LLVMBuildLoad2(self.llbuilder, ty, ptr, UNNAMED);
llvm::LLVMSetAlignment(load, align.bytes() as c_uint);

[antoyo]

llvm::LLVMSetAlignment(load, align.bytes() as c_uint);

The load alignment is somewhat confusing.
I don't remember for sure, but I believe the LLVMSetAlignment function applies to the load instruction and not the location where the value is loaded.
Since your fix makes some test fails, I would tend to think that the code is correct as is, but I'm might be wrong.

In some code I compiled, the align argument (edit: in Builder::load) wasn't the same as the pointer's regular alignment, and gcc ended up dereferencing the wrong address.

Could you please share an example where this happens?

[sapir]

I'll try to make an example tomorrow.

[sapir]

ok I think I've got it now. Also, I have some code to reproduce the problem.

// repr(C) to prevent any field reordering.
#[repr(C)]
struct X {
    // This causes the struct to require 8 byte alignment.
    a: u64,
    // This field is aligned to 8 bytes because the struct is aligned to 8 bytes, and it's at offset 8.
    // The type itself only requires 4 bytes, though.
    b: u32,
    c: u32,
    d: u32,
}

#[inline(never)]
fn foo(b1: u32, b2: u32) {
    dbg!(b1);
    dbg!(b2);
}

#[inline(never)]
fn bar(x: &X) {
    // x.b loads the value from b with 8-byte alignment.

    // The load result also has 8-byte alignment, so when being passed to foo,
    // the second argument gets aligned to 8-bytes and is passed at argument offset 8
    // as if it were the third 4-byte argument.

    foo(x.b, x.b);
}

fn main() {
    let x = X {
        a: 1,
        b: 2,
        c: 3,
        d: 4,
    };
    bar(&x);
}

So I think maybe the alignment shouldn't be applied to the load result.

[antoyo]

Can you explain what's the expected result with this code?

I get:

[src/main.rs:15:5] b1 = 2
[src/main.rs:16:5] b2 = 2

which seems good.

[sapir]

I'm getting 2 on the first line and some random value on the second.

[antoyo]

Which architecture are you on?

[sapir]

I was trying mips-unknown-linux-gnu. Since (edit: after) you asked, I also tried some other architectures including m68k but so far it hasn't happened anywhere except the mips one.

Also, it only happens with --release (I mean for cargo, e.g. y.sh cargo build --release). Without the --release, the argument gets passed fine.

[antoyo]

Do you know if it would be easy to add a test for this fix?

[sapir]

I think a test for the behavior might not be perfectly reliable because the bug could accidentally be optimized out, but I'll try. I think maybe with a higher field alignment it'll happen on m68k, too.

[antoyo]

Did you want to add a test or would you want to merge this as is?

[sapir]

I'd like to try to add a test that works on the architectures currently being tested in CI.

[antoyo]

Ok, please tell me when it's ready!
Thanks a lot for your work!