| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently various expressions handle this differently, and now we
consistently follow this rules:
---
For all non-control-flow value-returning instructions, if a type of an
expression is unreachable, we emit an unreachable and don't emit the
instruction itself. If we don't emit an unreachable, instructions that
follow can have validation failure in wasm binary format. For example:
```
[unreachable] (f32.add
[unreachable] (i32.eqz
[unreachable] (unreachable)
)
...
)
```
This is a valid prgram in binaryen IR, because the unreachable type
propagates out of an expression, making both i32.eqz and f32.add
unreachable. But in binary format, this becomes:
```
unreachable
i32.eqz
f32.add ;; validation failure; it expects f32 but takes an i32!
```
And here f32.add causes validation failure in wasm validation. So in this
case we add an unreachable to prevent following instructions to consume
the current value (here i32.eqz).
In actual tests, I used `global.get` to an f32 global, which does not
return a value, instead of `f32.add`, because `f32.add` itself will not
be emitted if one of argument is unreachable.
---
So the changes are:
- For instructions that don't return a value, removes unreachable
emitting code if it exists.
- Add the unreachable emitting code for value-returning instructions if
there isn't one.
- Check for unreachability only once after emitting all children for
atomic instructions. Currently only atomic instructions check
unreachability after visiting each children and bail out right after,
which is valid, but not consistent with others.
- Don't emit an extra unreachable after a return (and return_call). I
guess it is unnecessary.
|
| |
|
|
|
|
|
|
|
| |
When a memory instruction's type is unreachable, i.e., one of its
child expressions is unreachable, the instruction will be printed like
`unreachable.load`, which is invalid text format.
This prints unreachable prefix instruction types as `i32` to just make
them pass the parser. It is OK because they are not reachable anyway.
Also this removes printing of `?` in atomic.rmw instruction printing.
|
| |
|
|
|
|
|
|
|
| |
The blacklist means "functions here are to be ignored and not instrumented, we can assume they never unwind." The whitelist means "only these functions, and no others, can unwind." I had hoped such lists would not be necessary, since Asyncify's overhead is much smaller than the old Asyncify and Emterpreter, but as projects have noticed, the overhead to size and speed is still significant. The lists give power users a way to reduce any unnecessary overhead.
A slightly tricky thing is escaping of names: we escape names from the names section (see #2261 #1646). The lists arrive in human-readable format, so we escape them before comparing to the internal escaped names. To enable that I refactored wasm-binary a little bit to provide the escaping logic, cc @yurydelendik
If both lists are specified, an error is shown (since that is meaningless). If a name appears in a list that is not in the module, we show a warning, which will hopefully help people debug typos etc. I had hoped to make this an error, but the problem is that due to inlining etc. a single list will not always work for both unoptimized and optimized builds (a function may vanish when optimizing, due to duplicate function elimination or inlining).
Fixes #2218.
|
| |
|
|
|
|
|
| |
This adds
- `push`/`pop` support for other types: v128 and exnref
- `push`/`pop` support for binaryen.js
Because binaryen.js follows Binaryen's AST structure, without `pop` in
binaryen.js, EH instructions cannot be represented in binaryen.js.
|
| |
|
|
| |
Before I disallowed events with no values, but spec does not say
anything about it, so I think that restriction is not necessary.
|
| |
|
|
| |
Adds tail call support to fuzzer and makes small changes to handle return calls in multiple utilities and passes. Makes larger changes to DAE and inlining passes to properly handle tail calls.
|
| |
|
|
|
|
|
| |
The new flag indicates whether main reads the argc/argv parameters. If it does not, we can avoid emitting code to generate those arguments in the JS, which is not trivial in small programs - it requires some string conversion code.
Nicely the existing test inputs were enough for testing this (see outputs).
This depends on an emscripten change to land first, as emscripten.py asserts on metadata fields it doesn't recognize.
|
| |
|
|
|
|
|
|
|
|
|
| |
(global $g1 (mut i32) (i32.const 42))
(global $g2 i32 (global.get $g1))
can be optimized to
(global $g1 (mut i32) (i32.const 42))
(global $g2 i32 (i32.const 42))
even though $g1 is mutable - because it can't be mutated during module instantiation.
|
| |
|
|
| |
#2242 had exposed the bug that the `Trapper` pass was defining `walkFunction` when it should have been defining `doWalkFunction`.
|
| |
|
|
|
| |
(#2244)
This reverts commit 72c52ea7d4eb61b95cf8a5164947cb760fe42e9c, which was causing test failures after it merged.
|
| |
|
|
| |
This prevents those instructions from becoming invalid due to memory
packing optimizations and is also a code size win. Fixes #2227.
|
| | |
|
| | |
|
| |
|
|
|
|
|
| |
After some discussion this seems like a less confusing name: what the pass does is "asyncify" code, after all.
The one downside is the name overlaps with the old emscripten "Asyncify" utility, which we'll need to clarify in the docs there.
This keeps the old --bysyncify flag around for now, which is helpful for avoiding temporary breakage on CI as we move the emscripten side as well.
|
| |
|
|
| |
In WebAssembly/exception-handling#79 we agreed to rename `except_ref`
type to `exnref`.
|
| |
|
| |
We don't ever emit "use asm" anymore, so this similar annotation is not really useful, it just increases size.
|
| | |
|
| |
|
| |
This became noticeable after #2216 which led to some eqz eqz pairs in the test suite.
|
| |
|
|
|
|
|
| |
An if whose body is a br_if can be turned into a br_if of a combined condition (if side effects allow it).
The naive size in bytes is identical between the patterns, but the select may avoid a hardware branch, and also the select may be further optimized. On the benchmark suite this helps every single benchmark, but by quite small amounts (e.g. 100 bytes on sqlite, which is 1MB).
This was noticed in emscripten-core/emscripten#8941
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
We were passing bad value in --initial-stack-pointer which did not
include the STATIC_BUMP (since STATIC_BUMP is determinted by the output
of finalize).
If emscripten wants to set the stack pointer position it can do
so by calling the stackRestore() function at startup.
This argument will be removed completely once we stop passing it on the
emscripten side.
See https://github.com/emscripten-core/emscripten/issues/8905
|
| |
|
|
|
|
|
|
|
|
|
| |
Including parsing, printing, assembling, disassembling.
TODO:
- interpreting
- effects
- finalization and typing
- fuzzing
- JS/C API
|
| | |
|
| |
|
|
| |
Allow MemoryPacking to run when there are no passive segments, even if
bulk memory is enabled.
|
| |
|
|
|
| |
Fix and test mutable globals support, replace string literals with
constants, and add a pass to emit the target features section.
|
| |
|
|
|
|
|
| |
This is the first stage of adding support for stacky/multivaluey things. It adds new push/pop instructions, and so far just shows that they can be read and written, and that the optimizer doesn't do anything immediately wrong on them.
No fuzzer support, since there isn't a "correct" way to use these yet. The current test shows some "incorrect" usages of them, which is nice to see that we can parse/emit them, but we should replace them with proper usages of push/pop once we actually have those (see comments in the tests).
This should be enough to unblock exceptions (which needs a pop in try-catches). It is also a step towards multivalue (I added some docs about that), but most of multivalue is left to be done.
|
| |
|
| |
Previously we tried to export it if the memory was exported, even if growth was not on, which caused an error.
|
| |
|
|
|
| |
Add assertions on stack overflow in all 4 Bysyncify API calls (previously only 2 did it).
Also add a check that those assertions are hit.
|
| |
|
|
|
|
|
|
| |
Gets fuzzing support for Bysyncify working.
* Add the python to run the fuzzing on bysyncify.
* Add a JS script to load and run a testcase with bysyncify support. The code has all the runtime support for sleep/resume etc., which it does on calls to imports at random in a deterministic manner.
* Export memory from fuzzer so JS can access it.
* Fix tiny builder bug with makeExport.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
* Workaround for wasm2js output minification issue with emscripten
When using emscripten with -O2 and --memory-init-file 0, the
JS minification breaks on this function for memory initialization
setup, causing an exception to be thrown during module setup.
Moving from two 'var' declarations for the same variable to one
should avoid hitting this with no change in functionality (the
var gets hoisted anyway).
https://github.com/emscripten-core/emscripten/issues/8886
|
| |
|
| |
As decided in the recent in-person CG meeting.
|
| |
|
|
|
| |
We assigned it to a local, but didn't run maybeSkip on it. As a result, it was executed during rewinding, which broke restoring the saved value.
Found by the fuzzer.
|
| |
|
|
|
| |
We need memory in order to read and write rewinding info, so add it if the module didn't have any memory at all.
Found by the fuzzer.
|
| |
|
|
|
|
|
|
|
| |
(#2183)
This results in better code sizes on many testcases, sometimes much better. For example, on SQLite the 150K function has only 27 locals instead of 3,874 which it had before (!). This also reduces total code size on SQLite by 15%.
The key issue is that after instrumenting control flow we have a lot bigger live ranges.
This must be done rather carefully, as we need to introduce some temp locals early on (for breaking up ifs, for call return values, etc.).
|
| |
|
|
| |
This prevents RemoveImports from producing an invalid module that
references functions that no longer exist.
|
| |
|
|
|
| |
Those functions are assumed to be part of the runtime. Instrumenting them would mean nothing can work.
With this fix, bysyncify is useful with pure wasm, and not just through imports.
|
| |
|
| |
ignore-imports makes it not assume that any import may unwind/rewind the stack. ignore-indirect makes it not assume any indirect call can reach an unwind/rewind (which means, it assumes there is not an indirect call on the stack while unwinding).
|
| |
|
| |
Add a method to note the stopping of an unwind. This is enough to implement coroutines. Includes an example of coroutine usage in the test suite.
|
| |
|
|
|
|
|
|
|
| |
This adds a new pass, Bysyncify, which transforms code to allow unwind and rewinding the call stack and local state. This allows things like coroutines, turning synchronous code asynchronous, etc.
The new pass file itself has a large comment on top with docs.
So far the tests here seem to show this works, but this hasn't been tested heavily yet. My next step is to hook this up to emscripten as a replacement for asyncify/emterpreter, see emscripten-core/emscripten#8561
Note that this is completely usable by itself, so it could be useful for any language that needs coroutines etc., and not just ones using LLVM and/or emscripten. See docs on the ABI in the pass source.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixes in Relooper merge consecutive blocks:
Entry block getting removed when it is part of a loop:
bb1->AddBranchTo(bb2, nullptr);
bb1->AddBranchTo(bb3, ...);
bb2->AddBranchTo(bb1, nullptr);
bb3->AddBranchTo(bb4, nullptr);
relooper.AddBlock(bb1);
relooper.AddBlock(bb2);
relooper.AddBlock(bb3);
relooper.AddBlock(bb4);
relooper.Calculate(bb1);
Branches memory leak
|
| |
|
|
|
| |
This prevents the optimizer from producing v128.const instructions,
which are not supported by V8 at this time.
|
| |
|
|
|
|
|
| |
This should be small enough to work in a 512K stack on Linux, which may then be small enough to work on all common OSes.
I had to update some spec tests which actually did more recursive calls, but I don't think the change reduces any relevant amount of test coverage.
This may fix the Mac bot finally, as with this it passes for me on the stack size I think Macs have by default.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This adds support for the event and the event section, as specified in
https://github.com/WebAssembly/exception-handling/blob/master/proposals/Exceptions.md#changes-to-the-binary-model.
Wasm events are features that suspend the current execution and transfer
the control flow to a corresponding handler. Currently the only
supported event kind is exceptions.
For events, this includes support for
- Binary file reading/writing
- Wast file reading/writing
- Binaryen.js API
- Fuzzer
- Validation
- Metadce
- Passes: metrics, minify-imports-and-exports,
remove-unused-module-elements
|
| |
|
|
|
|
|
|
|
|
| |
* work
* fix
* fix
* format
|
| |
|
|
| |
This adds `Features.MVP` and `Features.All` to binaryen.js and make test
cases use it.
|
| |
|
|
|
| |
This pattern-matches towers of blocks + a br_table into a JS switch. This is much smaller in code size and also avoids heavy nesting that can exceed the recursion limits of JS parsers.
This is not enough yet, because it pattern-matches very specifically. In reality, switches can look slightly different. Followup PRs will extend this. For now, this passes the test suite (what passed before - not including the massive-switch tests) + fuzzing so it's a good start.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Refactored & fixed typeuse parsing rules so now the rules more closely
follow the spec. There have been multiple parsing rules that were
different in subtle ways, which are supposed to be the same according
to the spec.
- Duplicate types, i.e., types with the same signature, in the type
section are allowed as long as they don't have the same given name.
If a name is given, we use it; if type name is not given, we
generate one in the form of `$FUNCSIG$` + signature string. If the
same generated name already exists in the type section, we append
`_` at the end. This causes most of the changes in the autogenerated
type names in test outputs.
- A typeuse has to be in the order of (type) -> (param) -> (result),
if more than one of them exist. In case of function definitions,
(local) has to be after all of these. Fixed some test cases that
violate this rule.
- When only (param)/(result) are given, its type will be the type with
the smallest existing type index whose parameter and result are the
same. If there's no such type, a new type will be created and
inserted.
- Added a test case `duplicate_types.wast` to test type namings for
duplicate types.
- Refactored `parseFunction` function.
- Add more overrides to helper functions: `getSig` and
`ensureFunctionType`.
|
| |
|
|
|
| |
We have `getFunction`, but not `getGlobal` because its name clashed with
APIs for the deprecated instruction `get_global`. Now we have reflected
instruction renaming in code, we can add it for consistency.
|
| | |
|
| |
|
|
|
| |
(#2106)
Fixes #2103.
|
