Bug on non-power-of-2 SIMD vectors

[simnalamburt]

I think I found a bug with non-power-of-2 SIMD vectors. They're randomly turned into zero values when I store them on Vec

http://is.gd/LfOHbj

#[simd]
#[deriving(Show)]
struct f32x3(f32, f32, f32);

fn main() {
    let mut points = Vec::new();

    for _ in 0i..10 {
        points.push(f32x3(1.0, 1.0, 1.0));
    }

    for point in points.iter() {
        println!("{}", point);
    }
}

Expected result:

f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)

What we get:

f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(0, 0, 0)
f32x3(1, 1, 1)
f32x3(1, 1, 1)
f32x3(0, 0, 0)
f32x3(0, 0, 0)
f32x3(1, 1, 1)
f32x3(1, 1, 1)

Tested with:

rustc 0.13.0-nightly (39d740266 2015-01-01 15:51:08 +0000)
host: x86_64-apple-darwin

rustc 0.13.0-nightly (39d740266 2015-01-01 15:51:08 +0000)
host: x86_64-unknown-linux-gnu (Ubuntu 14.04.1 LTS)

[lifthrasiir]

http://is.gd/wRJF3h

As you can see, this is a fault of Vec assuming size_of::<T>() >= min_align_of::<T>() (while for T = f32x3 this is 12 and 16 respectively). More accurately, for T = f32x3, every growth allocates cap * 12 bytes while it actually needs cap * 16 bytes; any element after len >= cap * 3 / 4 will overflow the memory, and they will be omitted at the next reallocation.

[Gankra]

Ack!

[Gankra]

CC @thestinger

[simnalamburt]

I've got through some little experiment and I guess there's something wrong with size_of::<T>() for f32x3 and other non-power-of-2 SIMD types.

                     | size_of | min_align_of | align_of | size_of >= min_align_of
---------------------+---------+--------------+----------+-------------------------
 f32x3               |      12 |           16 |       16 | false
 f32x3_no_c          |      12 |           16 |       16 | false
 f32x3_no_simd       |      12 |            4 |        8 | true

                     | size_of | min_align_of | align_of | size_of >= min_align_of
---------------------+---------+--------------+----------+-------------------------
 u8                  |       1 |            1 |        1 | true
 OneByte             |       1 |            1 |        8 | true
 u16                 |       2 |            2 |        2 | true
 TwoByte             |       2 |            2 |        8 | true
 (u8, u16)           |       4 |            2 |        8 | true
 (u16, u8)           |       4 |            2 |        8 | true
 ThreeByte           |       4 |            2 |        8 | true
 (u32, u8)           |       8 |            4 |        8 | true
 FiveByte            |       8 |            4 |        8 | true
 (f32x3, u8)         |      32 |           16 |       16 | true
 (f32x3_no_simd, u8) |      16 |            4 |        8 | true

I think now we've got 2 available options.

1. Make size_of T >= min_align_of T always true even for non-power-of-2 SIMD types like f32x3.
2. Make Vec<T> function properly even for size_of T < min_align_of T situation.

I'm pretty new with Rust. Someone pleaes make me an advise!

[Aatch]

I'm not sure what the exact behaviour should be. Part of the problem is that non-power-of-two SIMD types don't exist on most platforms, so it gets widened to whatever it needs to be. However, an f32x3 is still only 12 bytes of actual data.

I think that Vec should handle the case anyway. Whether the case should come up... I don't really know.

[Diggsey]

Given that sizeof (f32x3, u32) == 32, it seems like rust already adds padding for SIMD types, so it's simply size_of that's broken, and should include the padding.

At the same time, Vec should probably assert that size_of T >= min_align_of T to catch any errors if in the future there are any types where their size is less than their alignment.

[Gankra]

Nominating because the required change may be a back-compat change to some primitive semantics.

[Aatch]

Actually, thinking about it some more. I believe that the current behaviour is correct. size_of doesn't account for padding, because padding is situation-dependent: (i8, u64) has 7 bytes of padding after the i8, but that doesn't stop i8 from being 1 byte in size. Just because the type gets padding, doesn't mean it needs to change size.

The whole point of having a different alignment to size is to cover these kinds of cases, where the number of usable bytes is not a nice integer multiple of the alignment.

[Diggsey]

@Aatch In that situation, the extra padding is added to meet the alignment requirements of "u64", not the "i8". In the SIMD example, the extra padding is part of the SIMD type.
ie. for (f32x3, u32) there are 4 bytes of padding between the fields, even though u32 only requires alignment to 4 bytes.

Either the compiler should not add any padding and sizeof (f32x3, u32) == 16, and any accesses take care to preserve the u32 or as is more usual, the SIMD type should include it's padding.

edit:
To clarify, types do include their padding in size_of, which is why size_of (u64, i8) == 16 as opposed to 9

[Diggsey]

Amusingly, struct Test(f32x3); has a size of 16, whereas f32x3 on it's own has a size of 12, which really shouldn't happen.

The issue is caused by size_of being implemented as a call to this function:

rust/src/librustc_trans/trans/machine.rs

Line 52 in 618bd5d

pub fn llsize_of_real(cx: &CrateContext, ty: Type) -> llsize {

The comment above the function is wrong: LLVM does include padding when calculating the size of types, but only for structs (not for vectorised types), which makes that function fairly useless and definitely not the preferred option as the comment suggests. llsize_of should be preferred instead which always includes padding, so is at least consistent.

[huonw]

Nominating because the required change may be a back-compat change to some primitive semantics.

NB. #[simd] and std::simd are experimental and are highly unlikely to be stable at 1.0.

[simnalamburt]

Current librustc works like below:

size_of::<T>() = ceil(LLVMSizeOfTypeInBits() / 8)
min_align_of::<T>() = LLVMABIAlignmentOfType()

and I found that current LLVM doesn't have any guarantee between LLVMSizeOfTypeInBits and LLVMABIAlignmentOfType. Instead, for SIMD vector types, LLVMABIAlignmentOfType is guaranteed to be greater or equal than LLVMSizeOfTypeInBits / 8. Look at this.

For example, LLVMSizeOfTypeInBits and LLVMABIAlignmentOfType of 80-bit floating point format is 80 and 16. I know that Rust doesn't have f80 but if it does, size_of::<f80>() will be 10 and min_align_of::<f80>() will be 16. Which means Vec<f80> would have same problem.

Now I see 2 options here

1. Make size_of::<T>() call llsize_of, not llsize_of_real (as @Diggsey said)
2. Make Vec<T> function properly even for size_of T < min_align_of T situation.

References

• size_of → llsize_of_real → LLVMSizeOfTypeInBits → llvm::DataLayout::getTypeSizeInBits
• llsize_of → llsize_of_alloc → LLVMABISizeOfType → llvm::DataLayout::getTypeAllocSize
• min_align_of → align_of → llalign_of_min → LLVMABIAlignmentOfType → llvm::DataLayout::getABITypeAlignment → llvm::DataLayout::getAlignment → llvm::DataLayout::getAlignmentInfo

[Gankra]

@huonw I meant more-so the size_of semantics, but simd semantics might also be affected.