This was previously using the following condition to decide that a block
hadn't been visited yet: either dest_offset is non-0 or the block isn't
the entry block. Unfortunately, this didn't work when the first block
would be non-empty but wouldn't generate code at all.
Since the original code would do at least one pass over the entire code,
the first pass that determines initial EBB offsets is done separately,
without considering branch relaxation. This ensures that all EBBs have
been visited and have correct initial offsets, and doesn't require a
special check to know whether an EBB has been visited or not.
* Support 32-bit build.
Signatures/functions/imports/exports etc are defined as varuint32 in
the WebAssembly specification so use u32 rather than u64.
Decrease the static memory constants for 32-bit addressing mode so
that they fit within 32-bit memory constraints.
Conditionalize cmake compile of SignalHandlers.cpp so that -m32 is
passed when building 32-bit.
Add a no-op match for Reloc::X86CallPCRel4 during linking. This is
probably the wrong thing, but it allows the tests to pass. Using the
same logic from the Reloc::X86PCRel4 case did not work.
Major API-incompatible changes include:
- Introduce TrapCode::UnreachableCodeReached, used for unreachable in wasm.
- cranelift-wasm's `declare_signature` now takes its signature by value
- cranelift-wasm's `declare_table_elements` `elems` parameter now takes a boxed slice
- Remove cranelift-wasm's `ModuleEnvironment`'s `get_signature`,
`get_num_func_imports`, `get_func_type`, and `get_global`.
This adds a feature which allows one to look up an export by name
without knowing what module it's in -- `lookup_global_export` on an
`InstanceContents`.
The main expected use for this is to support APIs where module A
imports a function from module B, and module B needs to access module
A's memory. B can't import it from A in the normal way, because that
would create a dependency cycle. So for now, allow B to look up A's
exported memory dynamically with `lookup_global_export`.
In the future, with reference types and possibly host bindings, we'll be
able to pass references to memory as arguments, which will obviate the
need for this mechanism.
We are facing peculiar Windows-only regressions in build times in
https://bugzilla.mozilla.org/show_bug.cgi?id=1506511 and while the build
times might just be Windows being slow, putting directories in
`rerun-if-changed` might also be causing problems. The build only
depends on the files, anyway, so let's just say that.
With Rust 2018 Edition, the `mod std` trick to alias `core` names to
`std` no longer works, so switch to just having the code use `core`
explicitly.
So instead, switch to just using `core::*` for things that in core.
This is more consistent with other Rust no_std code. And it allows
us to enable `no_std` mode unconditionally in the crates that support
it, which makes testing a little easier.
There actually three cases:
- For things in std and also in core, like `cmp`: Just use them via
`core::*`.
- For things in std and also in alloc, like `Vec`: Import alloc as std, as
use them from std. This allows them to work on both stable (which
doesn't provide alloc, but we don't support no_std mode anyway) and
nightly.
- For HashMap and similar which are not in core or alloc, import them in
the top-level lib.rs files from either std or the third-party hashmap_core
crate, and then have the code use super::hashmap_core.
Also, no_std support continues to be "best effort" at this time and not
something most people need to be testing.
