Add caching support for Singlepass backend.

[losfair]

This PR adds caching support for the Singlepass backend.

• Implementation
• AArch64 test

[losfair]

bors try

[bjfish]

Please activate this test:

wasmer/lib/runtime/src/cache.rs

Line 132 in dfc7163

#[cfg(all(test, not(feature = "singlepass")))]

[losfair]

bors try-

[losfair]

bors try

[nlewycky]

Question for you, how does this handle the code generated by Operator::F32Max which contains a pointer to the wasmer read-only data? Something like this:

                    static CANONICAL_NAN: u128 = 0x7FC0_0000;
                    // load float canonical nan
                    a.emit_mov(
                        Size::S64,
                        Location::Imm64((&CANONICAL_NAN as *const u128) as u64),
                        Location::GPR(tmpg1),
                    );

We can't assume the wasmer .rodata will be in the same place in memory each time, due to ASLR.

[nlewycky]

Is this the absolute offset in the file of the function, or is this the number of bytes between the function and the trampoline? Also does this mean that the order in the on-disk file is <function1, trampoline1>, <function2, trampoline2>, ...?

[losfair]

The absolute offset.

It is <trampoline1, function1>, <trampoline2, function2>, ... , though the trampolines only exist on AArch64.

[nlewycky]

Thanks! Could you clarify this in the comments? Maybe at the top with:

/// On-disk cache format.
/// Offsets are relative to the start of the file.
/// Trampolines are only present on AArch64.

[losfair]

@nlewycky Didn't notice that we now have references as absolute addresses to outside data in the generated code - this itself is not expected and should be fixed.

[bors]

try

Build succeeded

• wasmerio.wasmer

[nlewycky]

@nlewycky Didn't notice that we now have references as absolute addresses to outside data in the generated code - this itself is not expected and should be fixed.

OK. Unfortunately x86-64 doesn't have much of a way to put a constant into an xmm register except by going through memory. Maybe emit a jmp forward, write the constant bytes into the instruction stream, the load them as constant offset before %rip?

[bjfish]

This could also be updated: https://github.com/wasmerio/wasmer/blob/master/docs/feature_matrix.md

[losfair]

Seems that we are already using a temporary GPR to keep the address of the constant. Since the values are at most 64 bits wide, why not do mov imm{32,64}, gpr and then mov gpr, xmm directly? (updated code with this approach, waiting on tests)

[losfair]

bors try

[bors]

try

Build succeeded

• wasmerio.wasmer

[nlewycky]

Thanks! Could you clarify this in the comments? Maybe at the top with:

/// On-disk cache format.
/// Offsets are relative to the start of the file.
/// Trampolines are only present on AArch64.

[losfair]

bors r+

[bors]

Build succeeded

• wasmerio.wasmer

[ethanfrey]

Awesome! Looking forward to trying this out.