This is a followup to af2a952aabd82cf401cc664d0262b139ff92d86b. It
teaches the spilling pass to use the is_ghost() property to test whether
to visit instructions. This fixes a bug handling multiple return values
with fallthrough_return.
It was the caller's responsibility to call TargetIsa::check() before;
now one can't manipulate a TargetIsa without calling the
TargetIsaBuilder::finish() method, which is less error-prone and more in
line with what's coming for other things we're going to generate in the
meta crate.
Also splits the construction of a RegClass in two parts: a prototype is
made first when declaring the RegClass, and missing bits are filled in
when adding it to the TargetIsa(Builder). This avoids an awkward passing
of the isa to the RegClass ctor.
Add an explicit "is_ghost" property to selected instructions, and use
that to determine whether reload and coloring should visit instructions.
This allows them to visit fallthrough_return instructions and insert
fills and register moves as needed.
When one value is used multiple times for separate return values, we
need to copy it to produce a new value, so that each value can be
allocated a different register.
* Introduce a `TargetFrontendConfig` type.
`TargetFrontendConfig` is information specific to the target which is
provided to frontends to allow them to produce Cranelift IR for the
target. Currently this includes the pointer size and the default calling
convention.
The default calling convention is now inferred from the target, rather
than being a setting. cranelift-native is now just a provider of target
information, rather than also being a provider of settings, which gives
it a clearer role.
And instead of having cranelift-frontend routines require the whole
`TargetIsa`, just require the `TargetFrontendConfig`, and add a way to
get the `TargetFrontendConfig` from a `Module`.
Fixes#529.
Fixes#555.
* Rename size to base_size and introduce a compute_size function;
* Add infra to inspect in/outs registers when computing the size of an instruction;
* Remove the GPR_SAFE_DEREF and GPR_ZERO_DEREF_SAFE register classes on x86 (fixes#335);
* Put TargetIsa's emit_inst under a "testing_hooks" feature.
In practice, TargetIsa's emit_inst pulls in its own instantiation
of the target-specifi `emit_inst` functions, which can be quite
large, and LTO doesn't eliminate them because they're held live
by TargetIsa's vtable.
Fortunately, this function is only used by tests, so we can put
it behind a feature flag.
Fixes#530.
* Add comments for `emit_inst` to clarify its purpose.
* Move `return_at_end` out of Settings and into the wasm FuncEnvironment.
The `return_at_end` flag supports users that want to append a custom
epilogue to Cranelift-produced functions. It arranges for functions to
always return via a single return statement at the end, and users are
expected to remove this return to append their code.
This patch makes two changes:
- First, introduce a `fallthrough_return` instruction and use that
instead of adding a `return` at the end. That's simpler than having
users remove the `return` themselves.
- Second, move this setting out of the Settings and into the wasm
FuncEnvironment. This flag isn't something the code generator uses,
it's something that the wasm translator uses. The code generator
needs to preserve the property, however we can give the
`fallthrough_return` instruction properties to ensure this as needed,
such as marking it non-cloneable.
WebAssembly doesn't have non-dense jump tables, and higher-level users
are better served by the facilities in lib/frontend/src/switch.rs for
working with non-dense switches.
This eliminates the concept of "absent" jump table entries, which
were represented as "0" in the text format.
Also, jump table contents are now enclosed in `[` and `]`, so that
we can unambiguously display empty jump tables. Previously, empty jump
tables were displayed as if they had a single absent entry.
* Add encodings for i8 and i16 copy, spill, fill, ireduce.i8.i16
Also adds legalization for srem, irsub_imm, {u,s}extend.i16.i8
Fixes#477 cc #466
* Legalize popcnt, clz and ctz for i8 and i16
* Fix bug in call_memset
