* Reduce calls to `section_by_name` loading artifacts
Data is stored in binary artifacts as an ELF object and when loading an
artifact lots of calls are made to the `object` crate's
`section_by_name` method which ends up doing a linear search through the
list of sections for a particular name. To avoid doing this linear
search every time I've replaced this with one loop over the sections of
an object at the beginning when an object is loaded, or at least most of
the calls with this loop.
This isn't really a pressing issue today but some upcoming work I hope
to do for AOT-compiled components will be adding more sections to the
artifact so it seems best to keep the number of linear searches small
and avoided if possible.
* Fix an off-by-one
* Initial skeleton for Winch
This commit introduces the initial skeleton for Winch, the "baseline"
compiler.
This skeleton contains mostly setup code for the ISA, ABI, registers,
and compilation environment abstractions. It also includes the
calculation of function local slots.
As of this commit, the structure of these abstractions looks like the
following:
+------------------------+
| v
+----------+ +-----+ +-----------+-----+-----------------+
| Compiler | --> | ISA | --> | Registers | ABI | Compilation Env |
+----------+ +-----+ +-----------+-----+-----------------+
| ^
+------------------------------+
* Compilation environment will hold a reference to the function data
* Add basic documentation to the ABI trait
* Enable x86 and arm64 in cranelift-codegen
* Add reg_name function for x64
* Introduce the concept of a MacroAssembler and Assembler
This commit introduces the concept of a MacroAsesembler and
Assembler. The MacroAssembler trait will provide a high enough
interface across architectures so that each ISA implementation can use their own low-level
Assembler implementation to fulfill the interface. Each Assembler will
provide a 1-1 mapping to each ISA instruction.
As of this commit, only a partial debug implementation is provided for
the x64 Assembler.
* Add a newtype over PReg
Adds a newtype `Reg` over regalloc2::PReg; this ensures that Winch
will operate only on the concept of `Reg`. This change is temporary
until we have the necessary machinery to share a common Reg
abstraction via `cranelift_asm`
* Improvements to local calcuation
- Add `LocalSlot::addressed_from_sp`
- Use `u32` for local slot and local sizes calculation
* Add helper methods to ABIArg
Adds helper methods to retrieve register and type information from the argument
* Make locals_size public in frame
* Improve x64 register naming depending on size
* Add new methods to the masm interface
This commit introduces the ability for the MacroAssembler to reserve
stack space, get the address of a given local and perform a stack
store based on the concept of `Operand`s.
There are several motivating factors to introduce the concept of an
Operand:
- Make the translation between Winch and Cranelift easier;
- Make dispatching from the MacroAssembler to the underlying Assembler
- easier by minimizing the amount of functions that we need to define
- in order to satisfy the store/load combinations
This commit also introduces the concept of a memory address, which
essentially describes the addressing modes; as of this commit only one
addressing mode is supported. We'll also need to verify that this
structure will play nicely with arm64.
* Blank masm implementation for arm64
* Implementation of reserve_stack, local_address, store and fp_offset
for x64
* Implement function prologue and argument register spilling
* Add structopt and wat
* Fix debug instruction formatting
* Make TargetISA trait publicly accessible
* Modify the MacroAssembler finalize siganture to return a slice of strings
* Introduce a simple CLI for Winch
To be able to compile Wasm programs with Winch independently. Mostly
meant for testing / debugging
* Fix bug in x64 assembler mov_rm
* Remove unused import
* Move the stack slot calculation to the Frame
This commit moves the calculation of the stack slots to the frame
handler abstraction and also includes the calculation of the limits
for the function defined locals, which will be used to zero the locals
that are not associated to function arguments
* Add i32 and i64 constructors to local slots
* Introduce the concept of DefinedLocalsRange
This commit introduces `DefinedLocalsRange` to track the stack offset
at which the function-defined locals start and end; this is later used
to zero-out that stack region
* Add constructors for int and float registers
* Add a placeholder stack implementation
* Add a regset abstraction to track register availability
Adds a bit set abstraction to track register availability for register
allocation.
The bit set has no specific knowledge about physical registers, it
works on the register's hardware encoding as the source of truth.
Each RegSet is expected to be created with the universe of allocatable
registers per ISA when starting the compilation of a particular function.
* Add an abstraction over register and immediate
This is meant to be used as the source for stores.
* Add a way to zero local slots and an initial skeletion of regalloc
This commit introduces `zero_local_slots` to the MacroAssembler; which
ensures that function defined locals are zeroed out when starting the
function body.
The algorithm divides the defined function locals stack range
into 8 byte slots and stores a zero at each address. This process
relies on register allocation if the amount of slots that need to be
initialized is greater than 1. In such case, the next available
register is requested to the register set and it's used to store a 0,
which is then stored at every local slot
* Update to wasmparser 0.92
* Correctly track if the regset has registers available
* Add a result entry to the ABI signature
This commuit introduces ABIResult as part of the ABISignature;
this struct will track how function results are stored; initially it
will consiste of a single register that will be requested to the
register allocator at the end of the function; potentially causing a spill
* Move zero local slots and add more granular methods to the masm
This commit removes zeroing local slots from the MacroAssembler and
instead adds more granular methods to it (e.g `zero`, `add`).
This allows for better code sharing since most of the work done by the
algorithm for zeroing slots will be the same in all targets, except
for the binary emissions pieces, which is what gets delegated to the masm
* Use wasmparser's visitor API and add initial support for const and add
This commit adds initial support for the I32Const and I32
instructions; this involves adding a minimum for register
allocation. Note that some regalloc pieces are still incomplete, since
for the current set of supported instructions they are not needed.
* Make the ty field public in Local
* Add scratch_reg to the abi
* Add a method to get a particular local from the Frame
* Split the compilation environment abstraction
This commit splits the compilation environment into two more concise
abstractions:
1. CodeGen: the main abstraction for code generation
2. CodeGenContext: abstraction that shares the common pieces for
compilation; these pieces are shared between the code generator and
the register allocator
* Add `push` and `load` to the MacroAssembler
* Remove dead code warnings for unused paths
* Map ISA features to cranelift-codegen ISA features
* Apply formatting
* Fix Cargo.toml after a bad rebase
* Add component-compiler feature
* Use clap instead of structopt
* Add winch to publish.rs script
* Minor formatting
* Add tests to RegSet and fix two bugs when freeing and checking for
register availability
* Add tests to Stack
* Free source register after a non-constant i32 add
* Improve comments
- Remove unneeded comments
- And improve some of the TODO items
* Update default features
* Drop the ABI generic param and pass the word_size information directly
To avoid dealing with dead code warnings this commit passes the word
size information directly, since it's the only piece of information
needed from the ABI by Codegen until now
* Remove dead code
This piece of code will be put back once we start integrating Winch
with Wasmtime
* Remove unused enum variant
This variant doesn't get constructed; it should be added back once a
backend is added and not enabled by default or when Winch gets
integrated into Wasmtime
* Fix unused code in regset tests
* Update spec testsuite
* Switch the visitor pattern for a simpler operator match
This commit removes the usage of wasmparser's visitor pattern and
instead defaults to a simpler operator matching approach. This removes
the complexity of having to define all the visitor trait functions at once.
* Use wasmparser's Visitor trait with a different macro strategy
This commit puts back wasmparser's Visitor trait, with a sigle;
simpler macro, only used for unsupported operators.
* Restructure Winch
This commit restuructures Winch's parts. It divides the initial
approach into three main crates: `winch-codegen`,`wasmtime-winch` and `winch-tools`.
`wasmtime-winch` is reponsible for the Wasmtime-Winch integration.
`winch-codegen` is solely responsible for code generation.
`winch-tools` is CLI tool to compile Wasm programs, mainly for testing purposes.
* Refactor zero local slots
This commit moves the logic of zeroing local slots from the codegen
module into a method with a default implementation in the
MacroAssembler trait: `zero_mem_range`.
The refactored implementation is very similar to the previous
implementation with the only difference
that it doesn't allocates a general-purpose register; it instead uses
the register allocator to retrieve the scratch register and uses this
register to unroll the series of zero stores.
* Tie the codegen creation to the ISA ABI
This commit makes the relationship between the ISA ABI and the codegen
explicit. This allows us to pass down ABI-specific bit and pieces to
the codegeneration. In this case the only concrete piece that we need
is the ABI word size.
* Mark winch as publishable directory
* Revamp winch docs
This commit ensures that all the code comments in Winch are compliant
with the syle used in the rest of Wasmtime's codebase.
It also imptoves, generally the quality of the comments in some modules.
* Panic when using multi-value when the target is aarch64
Similar to x64, this commit ensures that the abi signature of the
current function doesn't use multi-value returns
* Document the usage of directives
* Use endianness instead of endianess in the ISA trait
* Introduce a three-argument form in the MacroAssembler
This commit introduces the usage of three-argument form for the
MacroAssembler interface. This allows for a natural mapping for
architectures like aarch64. In the case of x64, the implementation can
simply restrict the implementation asserting for equality in two of
the arguments of defaulting to a differnt set of instructions.
As of this commit, the implementation of `add` panics if the
destination and the first source arguments are not equal; internally
the x64 assembler implementation will ensure that all the allowed
combinations of `add` are satisfied. The reason for panicking and not
emitting a `mov` followed by an `add` for example is simply because register
allocation happens right before calling `add`, which ensures any
register-to-register moves, if needed.
This implementation will evolve in the future and this panic will be
lifted if needed.
* Improve the documentation for the MacroAssembler.
Documents the usage of three-arg form and the intention around the
high-level interface.
* Format comments in remaining modules
* Clean up Cargo.toml for winch pieces
This commit adds missing fields to each of Winch's Cargo.toml.
* Use `ModuleTranslation::get_types()` to derive the function type
* Assert that start range is always word-size aligned
* Make send and remove wrapper around WasiNnCtx·
This removes the wrapper around WasiNnCtx and no longer requires borrow_mut(). Once send/sync
changes in OpenVINO crate are merged in it will allow·use by frameworks that requires this trait.
* Bump openvino to compatible version.
* BackendExecutionContext should be Send and Sync
* Fix rust format issues.
* Update Cargo.lock for openvino
* Audit changes to openvino crates.
Wiggle generates code that instruments APIs with tracing code. This is
handy for diagnosing issues at runtime, but when inspecting the output
of Wiggle, it can make the generated code difficult for a human to
decipher. This change makes tracing a default but optional feature,
allowing users to avoid tracing code with commands like `cargo expand
--no-default-features`. This should be no change for current crates
depending on `wiggle`, `wiggle-macro`, and `wiggle-generate`.
review: add 'tracing' feature to wasi-common
review: switch to using macro configuration parsing
Co-authored-by: Andrew Brown <andrew.brown@intel.com>
* aarch64: Fix incorrect masking for small types on bmask
`bmask` was accidentally relying on the uppermost bits of the register
for small types.
This was found by fuzzgen, when it generated a shift left followed by
a bmask, the shift left shifted the bits out of the range of the input
type (i8), however these are not automatically cleared since they
remained inside the 32 bits of the register.
That caused issues when the bmask tried to compare the whole register
instead of just the bottom bits. The solution here is to mask the upper
bits for small types.
* aarch64: Emit 32bit cmp on bmask
This fixes an issue where bmask was accidentally comparing the
upper bits of the register by always using a 64bit cmp.
* riscv: Mask high bits in bmask
* riscv: Add compile tests for br{z,nz}
* riscv: Use shifts to mask 32bit values
This produces less code than the AND since that version needs to
load an immediate constant from memory.
* cranelift: Update test input to hexadecimal values
This makes it a bit more clear what is being tested.
* riscv: Use addiw for masking 32 bit values
Co-authored-by: Trevor Elliott <telliott@fastly.com>
* aarch64: Update bmask rule priority
Co-authored-by: Trevor Elliott <telliott@fastly.com>
Add a new instruction uadd_overflow_trap, which is a fused version of iadd_ifcout and trapif. Adding this instruction removes a dependency on the iflags type, and would allow us to move closer to removing it entirely.
The instruction is defined for the i32 and i64 types only, and is currently only used in the legalization of heap_addr.
* wiggle: no longer need to guard wasmtime integration behind a feature
this existed so we could use wiggle in lucet, but lucet is long EOL
* replace wiggle::Trap with wiggle::wasmtime_crate::Trap
* wiggle tests: unwrap traps because we cant assert_eq on them
* wasi-common: emit a wasmtime::Trap instead of a wiggle::Trap
formally add a dependency on wasmtime here to make it obvious, though
we do now have a transitive one via wiggle no matter what (and therefore
can get rid of the default-features=false on the wiggle dep)
* wasi-nn: use wasmtime::Trap instead of wiggle::Trap
there's no way the implementation of this func is actually
a good idea, it will panic the host process on any error,
but I'll ask @mtr to fix that
* wiggle test-helpers examples: fixes
* wasi-common cant cross compile to wasm32-unknown-emscripten anymore
this was originally for the WASI polyfill for web targets. Those days
are way behind us now.
* wasmtime wont compile for armv7-unknown-linux-gnueabihf either
This adds a new field `types` to `ModuleTranslation`, so that
consumers can have access to the module type information known after
validation has finished. This change is useful when consumers want to
have access to the type information in wasmparser's terms rather than
in wasmtime_environ's equivalent types (e.g. `WasmFuncType`).
Now that we aren't trying to do overlap checking in parallel, we can
fuse the loop that generates a list of rule pairs with the loop that
checks those pairs.
Removing the intermediate vector of pairs should save a little time and
memory. But it also means we're no longer borrowing from the `by_term`
HashMap, so we can use `into_iter` instead of `values` to move ownership
out of the map. That in turn means that we can use `into_iter` on each
vector of rules as well, which turns out to offer a slightly nicer idiom
for looping over all pairs, and also means we drop allocations as soon
as possible.
I also pushed grouping by priority earlier, so the O(n^2) all-pairs loop
runs over smaller lists. If we later find we want to know about overlaps
across different priorities, the definition of the map key is an easy
place to make that change.
* Cranelift: disable egraphs in fuzzing for now.
As per [this comment], with a few recent discussions it's become clear
that we want to refactor egraphs in a way that will subsume, or make
irrelevant, some of the recent fuzzbugs that have arisen (and likely
lead to others, which we'll want to fix!). Rather than chase these down
then refactor later, it probably makes sense not to spend the human time
or fuzzing time doing so. This PR turns off egraphs support in fuzzing
configurations for now, to be re-enabled later.
[this comment]: https://github.com/bytecodealliance/wasmtime/issues/5126#issuecomment-1291222515
* Disable in cranelift-fuzzgen as well.
In particular, this was found to happen in #5099 because a `Result`
projection node was not deduplicating across two separate `isplit`s that
created it. (This is a separate issue we should also fix; `needs_dedup`
is I think overly conservative because `Result` can project out a single
value from a pure or impure node, but the projection itself should be
treated like any other pure operator.)
In any case, if we have a value `v0` and two separate `Result { value:
v0, result: N, ty }` nodes, each of these will fill in the type `ty` for
the `N`th output of `v0`, and the second will idempotently overwrite the
first; we should loosen the assert so that it allows this case.
Fixes#5099. Fixes#5100.
* Cranelift: pass iterators to `ABIMachineSpec::compute_arg_locs`
Instead of slices. This gives us more flexibility to pass custom sequences
without needing to allocate a `Vec` to hold them and pass in as a slice.
* Cranelift: Avoid cloning `ir::Signature`s in `SigData::from_func_sig`
This avoids two heap allocations per signature that are unnecessary 99% of the
time.
* fix typo
* Simplify condition in `missing_struct_return`
As discussed in the 2022/10/19 meeting, this PR removes many of the branch and select instructions that used iflags, in favor if using brz/brnz and select in their place. Additionally, it reworks selectif_spectre_guard to take an i8 input instead of an iflags input.
For reference, the removed instructions are: br_icmp, brif, brff, trueif, trueff, and selectif.
Looks like #5091 wasn't enough and some of the APIs needed updating with
changes made in the meantime. I've updated the action here and
additionally made a separate change where the release isn't continually
created and deleted but instead left alone and only the tag is updated.
This should work for the `dev` release and avoids deleting/recreating on
each PR, sending out notifications for new releases.
* cranelift: Remove iconst.i128
* bugpoint: Report Changed when only one instruction is mutated
* cranelift: Fix egraph bxor rule
* cranelift: Remove some simple_preopt opts for i128
Eliminate a few remaining instances of non-SSA code.
Remove infrastructure previously used for non-SSA code emission.
Related cleanup around flags handling.
A local github action we have has been broken for about a month now
meaning that the `dev` tag isn't getting updated or getting new
releases. This appears to be due to the publication of new versions of
these dependencies which are running into issues using one another. I
think I've figured out versions that work and have added a
`package-lock.json` to ensure we keep using the same versions.
The `libfuzzer-sys` update in #5068 included some changes to the
`fuzz_target!` macro which caused a bare `run` function to be shadowed
by the macro-defined `run` function (changed in
rust-fuzz/libfuzzer#95) which meant that some of our fuzz targets were
infinite looping or stack overflowing as the same function was called
indefinitely. This renames the top-level `run` function to something
else in the meantime.
Using rayon adds a lot of dependencies to Cranelift. The total
unparallelized time the code that uses rayon takes is less than half a
second and it runs at compile time, so there is pretty much no benefit
to parallelizing it.
This can help rustc/llvm avoid bounds checks, but more importantly I will have
future changes here that remove indexing of params, and instead hand them out as
an iterator.
The `SigData::from_func_sig` constructor will already ensure that the struct
return pointer is returned, so this is a purely unnecessary call.
Note that this is not a performance speed up, since
`ensure_struct_return_ptr_is_returned` doesn't do any significant work if the
signature is already legalized.
This fixes#5086 by addressing two separate issues:
- The `ValueDataPacked::set_type()` helper had an embarrassing bitfield-manipulation bug that would mangle the rest of a `ValueDef` when setting its type. This is not normally used, only when the egraph elaboration fills in types after-the-fact on a multi-value node.
- The lowering rules for `isplit` on aarch64 and s390x were dispatching on the first output type, rather than the input type. When only the second output is used (as in the example in #5086), the first output type actually remains `INVALID` (and this is fine because it's never used).
Remove uses of reg_mod from the s390x backend. This required moving away from using r0/r1 as the result registers from a few different pseudo instructions, standardizing instead on r2/r3. That change was necessary as regalloc2 will not correctly allocate registers that aren't listed in the allocatable set, which r0/r1 are not.
Co-authored-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
Co-authored-by: Chris Fallin <chris@cfallin.org>
* Upgrade our github actions to "node16"
Each github actions run has a lot of warnings about using node12 so this
upgrades our repository to using node16. I'm hoping no other changes are
needed and I suspect other actions we're using are on node12 and will
need further updates, but this should help pin down what's remaining.
* Update `actions/checkout` workflow to `v3`
* Update to `actions/cache@v3`
* Update to `actions/upload-artifact@v3`
* Drop usage of `actions-rs/toolchain`
* Update to `actions/setup-python@v4`
* Update mdbook version
* fuzzgen: Test compiler flags
* cranelift: Generate `all()` function for all enum flags
This allows a user to iterate all flags that exist.
* fuzzgen: Minimize regalloc_checker compiles
* fuzzgen: Limit the amount of test case inputs
* fuzzgen: Add egraphs flag
It's finally here! 🥳
* cranelift: Add fuzzing comment to settings
* fuzzgen: Add riscv64
* fuzzgen: Unconditionally enable some flags
This commit fixes the `push-tag.yml` workflow to work with the new
`Cargo.toml` manifest since workspace inheritance was added. This
additionally fixes some warnings coming up on CI about our usage of
deprecated features on github actions.
* egraphs: a few miscellaneous compile-time optimizations.
These optimizations together are worth about a 2% compile-time
reduction, as measured on one core with spidermonkey.wasm as an input,
using `hyperfine` on `wasmtime compile`.
The changes included are:
- Some better pre-allocation (blockparams and side-effects concatenated
list vecs);
- Avoiding the indirection of storing list-of-types for every Pure and
Inst node, when almost all nodes produce only a single result;
instead, store arity and single type if it exists, and allow result
projection nodes to fill in types otherwise;
- Pack the `MemoryState` enum into one `u32` (this together with the
above removal of the type slice allows `Node` to
shrink from 48 bytes to 32 bytes);
- always-inline an accessor (`entry` on `CtxHash`) that wasn't
(`always(inline)` appears to be load-bearing, rather than just
`inline`);
- Split the update-analysis path into two hotpaths, one for the union
case and one for the new-node case (and the former can avoid
recomputing for the contained node when replacing a node with
node-and-child eclass entry).
* Review feedback.
* Fix test build.
* Fix to lowering when unused output with invalid type is present.
* func_wrap_async typechecks
* func call async
* instantiate_async
* fixes
* async engine creation for tests
* start adding a component model test for async
* fix wrong check for async support, factor out Instance::new_started to an unchecked impl
* tests: wibbles
* component::Linker::func_wrap: replace IntoComponentFunc with directly accepting a closure
We find that this makes the Linker::func_wrap type signature much easier
to read. The IntoComponentFunc abstraction was adding a lot of weight to
"splat" a set of arguments from a tuple of types into individual
arguments to the closure. Additionally, making the StoreContextMut
argument optional, or the Result<return> optional, wasn't very
worthwhile.
* Fixes for the new style of closure required by component::Linker::func_wrap
* future of result of return
* add Linker::instantiate_async and {Typed}Func::post_return_async
* fix fuzzing generator
* note optimisation opportunity
* simplify test
* Add egraphs option to Wasmtime config, and add it to fuzzing config generation.
This PR adds a wrapper method for Cranelift's `use_egraphs` setting to
Wasmtime's `Config`, named `cranelift_use_egraphs` analogously to its
existing `cranelift_opt_level`.
Eventually this should become a no-op as egraph-based optimization
becomes the default, but until then it makes sense to expose this as
another kind of optimization option.
This PR then adds the option to the `Arbitrary`-based config generation
for fuzzing, so compilation with egraphs will be fuzzed (on its own and
against other configurations and oracles).
* Don't use `NamedTempFile` on Windows
It looks like this prevents mmap-ing since the named temporary file
holds a `File` open which conflicts with the rights we're trying to open
the file for mmap-ing. Instead use a temporary directory to try to fix
this issue.
Co-authored-by: Alex Crichton <alex@alexcrichton.com>
* component::Linker::func_wrap: replace IntoComponentFunc with directly accepting a closure
We find that this makes the Linker::func_wrap type signature much easier
to read. The IntoComponentFunc abstraction was adding a lot of weight to
"splat" a set of arguments from a tuple of types into individual
arguments to the closure. Additionally, making the StoreContextMut
argument optional, or the Result<return> optional, wasn't very
worthwhile.
* Fixes for the new style of closure required by component::Linker::func_wrap
* fix fuzzing generator
Remove the boolean types from cranelift, and the associated instructions breduce, bextend, bconst, and bint. Standardize on using 1/0 for the return value from instructions that produce scalar boolean results, and -1/0 for boolean vector elements.
Fixes#3205
Co-authored-by: Afonso Bordado <afonso360@users.noreply.github.com>
Co-authored-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
Co-authored-by: Chris Fallin <chris@cfallin.org>
