Support 32-bit targets

[pepyakin]

To properly support them we need to setup CI for these targets

[eira-fransham]

I'm working on this. There's an issue with how sNaN is interpreted and I haven't found the cause yet

[eira-fransham]

Right, so I found the cause of the failures: when compiled for i686 (which is the 32-bit system I was using) the tests fail because the bit-level representation of floats is wrong. The IEEE754 spec allows NaNs to be quieted at basically any time but the WASM spec specifies that they may not be quietened when converting to and from i32, only when doing operations like multiplication. Unfortunely, on x87 LLVM will sometimes use the x87 float register stack to pass floats in and out of functions, using fld to push it on the callee and fstp to pop it on the caller. Unfortunately, these operations can and do mask the signalling bit. It works on x86_64 because the floating-point register stack was not inherited from x87 and it appears that it just uses %eap to return floats like it would return any other value.

The only solution I can see for this is to internally use integers to store floats, converting them to/from floats when doing * and /, but not when doing unary negation. In the long-term it might be better to use soft floats, since that would mean that we wouldn't have to disallow floating-point operations in the Polkadot WASM VM.

[sunfishcode]

Another possible approach is to build i686 builds with -C target-feature=+sse2 .

Beyond sNaNs, x87 also rounds some arithmetic results differently, due to double rounding, and WebAssembly's testsuite is designed to detect such differences. There are techniques for adjusting the rounding, though it's simpler if you can avoid having to.

[eira-fransham]

target-feature works for i686 but is not a generic solution. We should really only rely on behaviour specified in IEEE754 (which essentially means that we can only cast to "real" float when doing binary operations)

[sunfishcode]

If you're planning to support x87 floating-point, also be aware that x87's mul and div sometimes round differently on x87 than WebAssembly requires, even if you store the values between every operation, and even if you configure the x86 control register to 64-bit precision mode.

[eira-fransham]

If that is the case, then there are no tests for it so far.

[sunfishcode]

The spec testsuite's float_misc.wast has eg. this and this which should show it.

[eira-fransham]

float_misc succeeds on i686

[sunfishcode]

Other implementations fail on those tests when using with x87 arithmetic. I don't know what wasmi is doing differently that allows it to pass.

[sunfishcode]

With cargo test --target i686-unknown-linux-gnu --features opt-in-32bit --no-fail-fast, float_misc.wast passes for me. It appears Rust enables SSE2 by default.

With SSE2 disabled, via RUSTFLAGS="-C target-feature=-sse2" cargo test --target i686-unknown-linux-gnu --features opt-in-32bit --no-fail-fast, float_misc.wast fails for me, with a rounding difference:

Line 58:
thread 'spec::wasm_float_misc' panicked at 'assertion failed: `(left == right)`
  left: `F64(1.0)`,
 right: `F64(1.0000000000000002)`', tests/spec/run.rs:375:54

(this is just the first difference; other tests in float_misc.wast will also round differently too).

[pepyakin]

Guess we need to reopen then.

[eira-fransham]

I think enforcing SSE2 arithmetic is reasonable for now, but we'd probably want to emulate it down the line. Fixing floats on x87+SSE2 is a fix for the vast majority of 32-bit x86-family CPUs out there. There hasn't been an x86 CPU without SSE2 since 2001 as far as I can tell (apart from a 2013 microcontroller but I can't imagine that x86 microcontrollers are flying off the shelves)

[crsCR]

So, is there a way to sort this for a 32bit x86 CPU? Trying to set up on an old Acer Aspire one as only spare I can run 24/7 with Linux (Debian) on.

32 bit targets are not supported at the moment. You can use '32bit_opt_in' feature. See #43

[eira-fransham]

It should now work on post-2003 x86 CPUs (i.e. those with SSE2, if your CPU can build and run Rust programs with no custom configuration then this includes you). Maybe you need to run cargo update -p wasmi and if that doesn't work then we need to push a new version.

[crsCR]

Can you tell me where I need to run that update command from please. I have to CD to a directory with a Cargo.toml file or it won't run - I've tried a few places (lots of Cargo.toml files) but when compiling it still says wasmi V 0.1.3 and not the latest (0.2.0). It's an intel Atom N270 cpu - supports SSE2.

[eira-fransham]

cargo update will only update to compatible versions. Your Cargo.toml should specify a version for wasmi that is compatible with 0.2.0. Probably your Cargo.toml says wasmi = "0.1" or wasmi = "0.1.3", change that to wasmi = "0.2".

[crsCR]

You've got the Cargo.lock file in /Polkadot Master that refers to wasmi v0.1.3 several times then the Cargo.toml in /Polkadot Master/Polkadot/Parachain that mentions wasmi v0.1 under dependancies... (The Cargo.toml in /Polkadot Master doesn't mention wasmi.) So edit what and run the update from where?

[crsCR]

I got it downloading v.0.2.0 but can't stop the thing downloading v0.1.3 as well. :/

[crsCR]

OK, so I've edited all .toml and .lock files - and it looks like only wasmi v0.2.0 is going to compile - but now get this:

error[E0308]: mismatched types
  --> substrate/primitives/src/sandbox.rs:97:49
   |
97 |             TypedValue::F32(v_bits) => RuntimeValue::F32(f32::from_bits(v_bits as u32)),
   |                                                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ expected struct `nan_preserving_float::F32`, found f32
   |
   = note: expected type `nan_preserving_float::F32`
              found type `f32`

error[E0308]: mismatched types
  --> substrate/primitives/src/sandbox.rs:98:49
   |
98 |             TypedValue::F64(v_bits) => RuntimeValue::F64(f64::from_bits(v_bits as u64)),
   |                                                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ expected struct `nan_preserving_float::F64`, found f64
   |
   = note: expected type `nan_preserving_float::F64`
              found type `f64`

[pepyakin]

If you're playing with Polkadot, then you can easily replace this code with unimplemented! macro. I think this is the easiest way.

Eventually, we need to migrate Polkadot to wasmi 0.2.0 though.

[crsCR]

Meh - looks like I'll have to install a VM on my Windows 64bit desktop...