* x64: Add rudimentary support for some AVX instructions
I was poking around Spidermonkey's wasm backend and saw that the various
assembler functions used are all `v*`-prefixed which look like they're
intended for use with AVX instructions. I looked at Cranelift and it
currently doesn't have support for many AVX-based instructions, so I
figured I'd take a crack at it!
The support added here is a bit of a mishmash when viewed alone, but my
general goal was to take a single instruction from the SIMD proposal for
WebAssembly and migrate all of its component instructions to AVX. I, by
random chance, picked a pretty complicated instruction of `f32x4.min`.
This wasm instruction is implemented on x64 with 4 unique SSE
instructions and ended up being a pretty good candidate.
Further digging about AVX-vs-SSE shows that there should be two major
benefits to using AVX over SSE:
* Primarily AVX instructions largely use a three-operand form where two
input registers are operated with and an output register is also
specified. This is in contrast to SSE's predominant
one-register-is-input-but-also-output pattern. This should help free
up the register allocator a bit and additionally remove the need for
movement between registers.
* As #4767 notes the memory-based operations of VEX-encoded instructions
(aka AVX instructions) do not have strict alignment requirements which
means we would be able to sink loads and stores into individual
instructions instead of having separate instructions.
So I set out on my journey to implement the instructions used by
`f32x4.min`. The first few were fairly easy. The machinst backends are
already of the shape "take these inputs and compute the output" where
the x86 requirement of a register being both input and output is
postprocessed in. This means that the `inst.isle` creation helpers for
SSE instructions were already of the correct form to use AVX. I chose to
add new `rule` branches for the instruction creation helpers, for
example `x64_andnps`. The new `rule` conditionally only runs if AVX is
enabled and emits an AVX instruction instead of an SSE instruction for
achieving the same goal. This means that no lowerings of clif
instructions were modified, instead just new instructions are being
generated.
The VEX encoding was previously not heavily used in Cranelift. The only
current user are the FMA-style instructions that Cranelift has at this
time. These FMA instructions have one extra operand than `vandnps`, for
example, so I split the existing `XmmRmRVex` into a few more variants to
fit the shape of the instructions that needed generating for
`f32x4.min`. This was accompanied then with more AVX opcode definitions,
more emission support, etc.
Upon implementing all of this it turned out that the test suite was
failing on my machine due to the memory-operand encodings of VEX
instructions not being supported. I didn't explicitly add those in
myself but some preexisting RIP-relative addressing was leaking into the
new instructions with existing tests. I opted to go ahead and fill out
the memory addressing modes of VEX encoding to get the tests passing
again.
All-in-all this PR adds new instructions to the x64 backend for a number
of AVX instructions, updates 5 existing instruction producers to use AVX
instructions conditionally, implements VEX memory operands, and adds
some simple tests for the new output of `f32x4.min`. The existing
runtest for `f32x.min` caught a few intermediate bugs along the way and
I additionally added a plain `target x86_64` to that runtest to ensure
that it executes with and without AVX to test the various lowerings.
I'll also note that this, and future support, should be well-fuzzed
through Wasmtime's fuzzing which may explicitly disable AVX support
despite the machine having access to AVX, so non-AVX lowerings should be
well-tested into the future.
It's also worth mentioning that I am not an AVX or VEX or x64 expert.
Implementing the memory operand part for VEX was the hardest part of
this PR and while I think it should be good someone else should
definitely double-check me. Additionally I haven't added many
instructions to the x64 backend yet so I may have missed obvious places
to tests or such, so am happy to follow-up with anything to be more
thorough if necessary.
Finally I should note that this is just the tip of the iceberg when it
comes to AVX. My hope is to get some of the idioms sorted out to make it
easier for future PRs to add one-off instruction lowerings or such.
* Review feedback
I saw some PRs fail this step earlier today due to rate limits but it
ended up not failing the entire PR's CI due to it not being listed in
the final set of dependencies, so add it there.
* Refactor collect_branches_and_targets to not need a smallvec
Basic blocks are terminated by at most one branch instruction now, so we
can use that assumption in `collect_branches_and_targets` to return the
last instruction we saw instead.
* Review comments
This is a short-term fix to the same bug that #5800 is addressing
(#5796), but with less risk: it simply turns off GVN'ing of effectful
but idempotent ops. Because we have an upcoming release, and this is a
miscompile (albeit to do with trapping behavior), we would like to make
the simplest possible fix that avoids the bug, and backport it. I will
then rebase #5800 on top of a revert of this followed by the more
complete fix.
* Don't run LLDB tests on PRs
These take an extra minute or so, so only run them on the full test
suite of a merge instead of on all PRs as well.
* Add a test for the x64 isa files
This guarantees that if cranelift's x64 backend is modified that the
tests will be run on a PR, even if other backends were also modified.
These mostly only validate changes to `Cargo.lock` so skip these checks
by default on PRs which generally never need to trigger them. If
`Cargo.lock` changes, however, then run them for PRs.
GitHub recently made its merge queue feature available for use in public
repositories owned by organizations meaning that the Wasmtime repository
is a candidate for using this. GitHub's Merge Queue feature is a system
that's similar to Rust's bors integration where PRs are tested before
merging and only passing PRs are merged. This implements the "not rocket
science" rule where the `main` branch of Wasmtime, for example, is
always tested and passes CI. This is in contrast to our current
implementation of CI where PRs are merged when they pass their own CI,
but the code that was tested is not guaranteed to be the state of `main`
when the PR is merged, meaning that we're at risk now of a failing
`main` branch despite all merged PRs being green. While this has
happened with Wasmtime this is not a common occurrence, however.
The main motivation, instead, to use GitHub's Merge Queue feature is
that it will enable Wasmtime to greatly reduce the amount of CI running
on PRs themselves. Currently the full test suite runs on every push to
every PR, meaning that our workers on GitHub Actions are frequently
clogged throughout weekdays and PRs can take quite some time to come
back with a successful run. Through the use of a Merge Queue, however,
we're able to configure only a small handful of checks to run on PRs
while deferring the main body of checks to happening on the
merge-via-the-queue itself. This is hoped to free up capacity on CI and
overall improve CI times for Wasmtime and Cranelift developers.
The implementation of all of this required quite a lot of plumbing and
retooling of our CI. I've been testing this in an [external
repository][testrepo] and I think everything is working now. A list of
changes made in this PR are:
* The `build.yml` workflow is merged back into the `main.yml` workflow
as the original reason to split it out is not longer applicable (it'll
run on all merges). This was also done to fit in the dependency graph
of jobs of one workflow.
* Publication of the `gh-pages` branch, the `dev` tag artifacts, and
release artifacts have been moved to a separate
`publish-artifacts.yml` workflow. This workflow runs on all pushes to
`main` and all tags. This workflow no longer actually preforms any
builds, however, and relies on a merge queue or similar being used for
branches/tags where artifacts are downloaded from the workflow run to
be uploaded. For pushes to `main` this works because a merge queue is
run meaning that by the time the push happens all artifacts are ready.
For release branches this is handled by..
* The `push-tag.yml` workflow is subsumed by the `main.yml` workflow. CI
for a tag being pushed will upload artifacts to a release in GitHub,
meaning that all builds must finish first for the commit. The
`main.yml` workflow at the end now scans commits for the preexisting
magical marker and pushes a tag if necessary.
* CI is currently a flat list of "run all these jobs" and this is now
rearchitected to a "fan out" approach where some jobs run to determine
the next jobs to run which then get "joined" into a finish step. The
purpose for this is somewhat nuanced and this has implications for CI
runtime as well. The Merge Queue feature requires branches to be
protected with "these checks must pass" and then the same checks are
gates both to enter the merge queue as well as pass the merge queue.
The saving grace, however, is that a "skipped" check counts as
passing, meaning checks can be skipped on PRs but run to completion on
the merge queue. A problem with this though is the build matrix used
for tests where PRs want to only run one element of the build matrix
ideally but there's no means on GitHub Actions right now for the
skipped entries to show up as skipped easily (or not that I know of).
This means that the "join" step serves the purpose of being the single
gate for both PR and merge queue CI and there's just more inputs to it
for merge queue CI. The major consequence of this decision is that
GitHub's actions scheduling doesn't work out well here. Jobs are
scheduled in a FIFO order meaning that the job for "ok complete the CI
run" is queued up after everything else has completed, possibly
after lots of other CI requests in the middle for other PRs. The hope
here is that by using a merge queue we can keep CI relatively under
control and this won't affect merge times too much.
* All jobs in the `main.yml` workflow will not automatically cancel the
entire run if they fail. Previously this fail-fast behavior was only
part of the matrix runs (and just for that matrix), but this is
required to make the merge queue expedient. The gate of the merge
queue is the final "join" step which is only executed once all
dependencies have finished. This means, for example, that if rustfmt
fails quickly then the tests which take longer might run for quite
awhile before the join step reports failure, meaning that the PR sits
in the queue for longer than needed being tested when we know it's
already going to fail. By having all jobs cancel the run this means
that failures immediately bail out and mark the whole job as
cancelled.
* A new "determine" CI job was added to determine what CI actually needs
to run. This is a "choke point" which is scheduled at the start of CI
that quickly figures out what else needs to be run. This notably
indicates whether large swaths of ci (the `run-full` flag) like the
build matrix are executed. Additionally this dynamically calculates a
matrix of tests to run based on a new `./ci/build-test-matrix.js`
script. Various inputs are considered for this such as:
1. All pushes, meaning merge queue branches or release-branch merges,
will run full CI.
2. PRs to release branches will run full CI.
3. PRs to `main`, the most common, determine what to run based on
what's modified and what's in the commit message.
Some examples for (3) above are if modifications are made to
`cranelift/codegen/src/isa/*` then that corresponding builder is
executed on CI. If the `crates/c-api` directory is modified then the
CMake-based tests are run on PRs but are otherwise skipped.
Annotations in commit messages such as `prtest:*` can be used to
explicitly request testing.
Before this PR merges to `main` would perform two full runs of CI: one
on the PR itself and one on the merge to `main`. Note that the one as a
merge to `main` was quite frequently cancelled due to a merge happening
later. Additionally before this PR there was always the risk of a bad
merge where what was merged ended up creating a `main` that failed CI to
to a non-code-related merge conflict.
After this PR merges to `main` will perform one full run of CI, the one
as part of the merge queue. PRs themselves will perform one test job
most of the time otherwise. The `main` branch is additionally always
guaranteed to pass tests via the merge queue feature.
For release branches, before this PR merges would perform two full
builds - one for the PR and one for the merge. A third build was then
required for the release tag itself. This is now cut down to two full
builds, one for the PR and one for the merge. The reason for this is
that the merge queue feature currently can't be used for our
wildcard-based `release-*` branch protections. It is now possible,
however, to turn on required CI checks for the `release-*` branch PRs so
we can at least have a "hit the button and forget" strategy for merging
PRs now.
Note that this change to CI is not without its risks. The Merge Queue
feature is still in beta and is quite new for GitHub. One bug that
Trevor and I uncovered is that if a PR is being tested in the merge
queue and a contributor pushes to their PR then the PR isn't removed
from the merge queue but is instead merged when CI is successful, losing
the changes that the contributor pushed (what's merged is what was
tested). We suspect that GitHub will fix this, however.
Additionally though there's the risk that this may increase merge time
for PRs to Wasmtime in practice. The Merge Queue feature has the ability
to "batch" PRs together for a merge but this is only done if concurrent
builds are allowed. This means that if 5 PRs are batched together then 5
separate merges would be created for the stack of 5 PRs. If the CI for
all 5 merged together passes then everything is merged, otherwise a PR
is kicked out. We can't easily do this, however, since a major purpose
for the merge queue for us would be to cut down on usage of CI builders
meaning the max concurrency would be set to 1 meaning that only one PR
at a time will be merged. This means PRs may sit in the queue for awhile
since previously many `main`-based builds are cancelled due to
subsequent merges of other PRs, but now they must all run to 100%
completion.
[testrepo]: https://github.com/bytecodealliance/wasmtime-merge-queue-testing
Rework br_table to use BlockCall, allowing us to avoid adding new nodes during ssa construction to hold block arguments. Additionally, many places where we previously matched on InstructionData to extract branch destinations can be replaced with a use of branch_destination or branch_destination_mut.
This change adds the `wasmtime-wasi-threads` crate as a default crate
for the CLI application. This is no change for embedders of Wasmtime:
they would still have to include `wasmtime-wasi-threads` manually.
Enabling the crate by default in the CLI application has several
benefits, e.g., that it is simpler to experiment with and that it will
be part of more test runs (and thus bugs can be discovered more
quickly). Users will still have to add
`--wasi-modules=experimental-wasi-threads` to enable wasi-threads on the
command line.
As a follow-up to #5780, disassemble the regions identified by bb_starts, falling back on disassembling the whole buffer. This ensures that instructions like br_table that introduce a lot of constants don't throw off capstone for the remainder of the function.
---------
Co-authored-by: Jamey Sharp <jamey@minilop.net>
* cranelift: Add `adrp` encoding to AArch64 backend
* cranelift: Support GOT Symbol References in AArch64
* cranelift: Add MachO GOT relocations
* cranelift: Do not mark the GOT PageOffset12 MachO relocation as relative
This updates the signatures of the `xmm_rm_r` helper function and then
updates existing users and migrates other users to the helper now that
the type information is no longer required.
* Update world-selection in `bindgen!` macro
Inspired by bytecodealliance/wit-bindgen#494 specifying a world or
document to bindgen is now optional as it's inferred if there's only one
`default world` in a package's documents.
* Add cargo-vet entry
This commit fixes a panic related to type imports where an import of a
type didn't correctly declare the new type index on the Wasmtime side of
things. Additionally this plumbs more support throughout Wasmtime to
support type imports, namely that they do not need to be supplied
through a `Linker`. This additionally implements a feature where empty
instances, even transitively, do not need to be supplied by a Wasmtime
embedder. This means that instances which only have types, for example,
do not need to be supplied into a `Linker` since no runtime information
for them is required anyway.
Closes#5775
This works around a `rustc` bug where compiling with LTO
will sometimes strip out some of the trampoline entrypoint
symbols resulting in a linking failure.
* Add (bnot (bxor x y)) lowerings for s390x/aarch64
I originally thought that s390x's original lowering in #5709, but as was
rightfully pointed out `(bnot (bxor x y))` is equivalent to
`(bxor x (bnot y))` so the special lowering for one should apply as a
special lowering for the other. For the s390x and aarch64 backend that
have already have a fused lowering of the bxor/bnot add a lowering
additionally for the bnot/bxor combination.
* Add bnot(bxor(..)) tests for s390x 128-bit sizes
* Update wasm-tools crates
Pulls in a new component binary format which should hopefully be the
last update for awhile.
* Update cargo vet configuration
* Add support for WASI sockets to C API
Add support for WASI sockets in the C API by adding a new API to handle
preopening sockets for clients. This uses HashMap instead of Vec for
preopened sockets to identify if caller has called in more than once
with the same FD number. If so, then we return false so caller is given
hint that they are attempting to overwrite an already existing socket
FD.
* Apply suggestions from code review
Co-authored-by: Peter Huene <peter@huene.dev>
* s/stdlistener/listener/
---------
Co-authored-by: Peter Huene <peter@huene.dev>
As jump tables are used by at most one br_table instruction, inline their definition in those instructions instead of requiring them to be declared as function-level metadata.
This patch adds complete support for the `sub` and `add` WebAssembly instructions
for x64, and complete support for the `add` WebAssembly instruction for aarch64.
This patch also refactors how the binary operations get constructed within the
`VisitOperator` trait implementation. The refactor adds methods in the
`CodeGenContext` to abstract all the common steps to emit binary operations,
making this process less repetitive and less brittle (e.g. omitting to push the resulting value
to the stack, or omitting to free registers after used).
This patch also improves test coverage and refactors the filetests directory to make it
easier to add tests for other instructions.
* wasi-threads: fix import name
As @TerrorJack pointed out in #5484, that PR implements an older
name--`thread_spawn`. This change uses the now-official name from the
specification--`thread-spawn`.
* fix: update name in test
Move the storage for jump tables off of FunctionStencil and onto DataFlowGraph. This change is in service of #5731, making it easier to access the jump table data in the context of helpers like inst_values.
* Generalize `n ^ !n` optimization to more types
* Generalize `x & -1` optimization to more types
Also mark the `x & x` rewrite to `subsume`.
* Cranelift: Optimize x|!x and x&!x to constants
These cases are much like the existing x^!x rules.
Similar to when we exposed the DataFlowGraph::insts field through a restrictive newtype, expose DataFlowGraph::blocks through an interface that allows a restrictive set of operations. This field being public now allows us to avoid a rematch in ssa construction, and simplifies the implementation of adding a block argument to a block referenced by a br_table instruction.
We don't need to spend time going through the GVN map to dedup a
newly-constructed `iconst 0` when we already matched that value on the
left-hand side of these rules.
Also, mark these rules as subsuming any others since we can't do better
than reducing an expression to a constant.
