| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously only WalkerPasses had access to the `getPassRunner` and
`getPassOptions` methods. Move those methods to `Pass` so all passes can use
them. As a result, the `PassRunner` passed to `Pass::run` and
`Pass::runOnFunction` is no longer necessary, so remove it.
Also update `Pass::create` to return a unique_ptr, which is more efficient than
having it return a raw pointer only to have the `PassRunner` wrap that raw
pointer in a `unique_ptr`.
Delete the unused template `PassRunner::getLast()`, which looks like it was
intended to enable retrieving previous analyses and has been in the code base
since 2015 but is not implemented anywhere.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Merge similar functions that only differs constant values (like immediate
operand of const and call insts) by parameterization.
Performing this pass at post-link time can merge more functions across
objects. Inspired by Swift compiler's optimization which is derived from
LLVM's one:
https://github.com/apple/swift/blob/main/lib/LLVMPasses/LLVMMergeFunctions.cpp
https://github.com/llvm/llvm-project/blob/main/llvm/docs/MergeFunctions.rst
The basic ideas here are constant value parameterization and direct callee
parameterization by indirection.
Constant value parameterization is like below:
;; Before
(func $big-const-42 (result i32)
[[many instr 1]]
(i32.const 44)
[[many instr 2]]
)
(func $big-const-43 (result i32)
[[many instr 1]]
(i32.const 45)
[[many instr 2]]
)
;; After
(func $byn$mgfn-shared$big-const-42 (result i32)
[[many instr 1]]
(local.get $0) ;; parameterized!!
[[many instr 2]]
)
(func $big-const-42 (result i32)
(call $byn$mgfn-shared$big-const-42
(i32.const 42)
)
)
(func $big-const-43 (result i32)
(call $byn$mgfn-shared$big-const-42
(i32.const 43)
)
)
Direct callee parameterization is similar to the constant value parameterization,
but it parameterizes callee function i by ref.func instead. Therefore it is enabled
only when reference-types and typed-function-references features are enabled.
I saw 1 ~ 2 % reduction for SwiftWasm binary and Ruby's wasm port
using wasi-sdk, and 3 ~ 4.5% reduction for Unity WebGL binary when -Oz.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Technically this is not a new pass, but it is a rewrite almost from scratch.
Local Common Subexpression Elimination looks for repeated patterns,
stuff like this:
x = (a + b) + c
y = a + b
=>
temp = a + b
x = temp + c
y = temp
The old pass worked on flat IR, which is inefficient, and was overly
complicated because of that. The new pass uses a new algorithm that
I think is pretty simple, see the detailed comment at the top.
This keeps the pass enabled only in -O4, like before - right after
flattening the IR. That is to make this as minimal a change as possible.
Followups will enable the pass in the main pipeline, that is, we will
finally be able to run it by default. (Note that to make the pass work
well after flatten, an extra simplify-locals is added - the old pass used
to do part of simplify-locals internally, which was one source of
complexity. Even so, some of the -O4 tests have changes, due to
minor factors - they are just minor orderings etc., which can be
seen by inspecting the outputs before and after using e.g.
--metrics)
This plus some followup work leads to large wins on wasm GC output.
On j2cl there is a common pattern of repeated struct.gets, so common
that this pass removes 85% of all struct.gets, which makes the total
binary 15% smaller. However, on LLVM-emitted code the benefit is
minor, less than 1%.
|
| |
|
|
|
|
|
|
|
| |
When using nominal types, func.ref of two functions with identical signatures
but different HeapTypes will yield different types. To preserve these semantics,
Functions need to track their HeapTypes, not just their Signatures.
This PR replaces the Signature field in Function with a HeapType field and adds
new utility methods to make it almost as simple to update and query the function
HeapType as it was to update and query the Function Signature.
|
| |
|
|
|
|
|
|
|
| |
* Unifies internal hashing helpers to naturally integrate with std::hash
* Removes the previous custom implementation
* Computed hashes are now always size_t
* Introduces a hash_combine helper
* Fixes an overwritten partial hash in Relooper.cpp
|
| |
|
|
|
|
| |
This involves replacing `Literal::makeZero` with `Literal::makeZeroes`
and `Literal::makeSingleZero` and updating `isConstantExpression` to
handle constant tuples as well. Also makes `Literals` its own struct
and adds convenience methods on it.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Now that we have subtypes, we cannot reuse any local that contains the
same expression, because that local's type can be a supertype. For
example:
```
(local $0 anyref)
(local $1 nullref)
...
(local.set $0 (ref.null))
(local.set $1 (ref.null)) ;; cannot be replaced with (local.get $0)
```
This extends `usables` map's key to contain both `HashedExpression` and
the local's type, so we can get the right usable local in presence of
subtypes.
|
| |
|
|
|
|
|
|
|
|
| |
* Remove implicit conversion operators from Type
Now types must be explicitly converted to uint32_t with Type::getID or
to ValueType with Type::getVT. This fixes #2572 for switches that use
Type::getVT.
* getVT => getSingle
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Function signatures were previously redundantly stored on Function
objects as well as on FunctionType objects. These two signature
representations had to always be kept in sync, which was error-prone
and needlessly complex. This PR takes advantage of the new ability of
Type to represent multiple value types by consolidating function
signatures as a pair of Types (params and results) stored on the
Function object.
Since there are no longer module-global named function types,
significant changes had to be made to the printing and emitting of
function types, as well as their parsing and manipulation in various
passes.
The C and JS APIs and their tests also had to be updated to remove
named function types.
|
| |
|
| |
Applies the changes in #2065, and temprarily disables the hook since it's too slow to run on a change this large. We should re-enable it in a later commit.
|
| |
|
| |
Mass change to apply clang-format to everything. We are applying this in a PR by me so the (git) blame is all mine ;) but @aheejin did all the work to get clang-format set up and all the manual work to tidy up some things to make the output nicer in #2048
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously the relooper would do some optimizations when deciding when to use an if vs a switch, how to group blocks, etc. This PR adds an additional pre-optimization phase with some basic but useful simplify-cfg style passes,
* Skip empty blocks when they have just one exit.
* Merge exiting branches when they are equivalent.
* Canonicalize block contents to make such comparisons more useful.
* Turn a trivial one-target switch into a simple branch.
This can help in noticeable ways when running the rereloop pass, e.g. on LLVM wasm backend output.
Also:
* Binaryen C API changes to the relooper, which now gets a Module for its constructor. It needs it for the optimizations, as it may construct new nodes.
* Many relooper-fuzzer improvements.
* Clean up HashType usage.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This adds a new IR, "Stack IR". This represents wasm at a very low level, as a simple stream of instructions, basically the same as wasm's binary format. This is unlike Binaryen IR which is structured and in a tree format.
This gives some small wins on binary sizes, less than 1% in most cases, usually 0.25-0.50% or so. That's not much by itself, but looking forward this prepares us for multi-value, which we really need an IR like this to be able to optimize well. Also, it's possible there is more we can do already - currently there are just a few stack IR optimizations implemented,
DCE
local2stack - check if a set_local/get_local pair can be removed, which keeps the set's value on the stack, which if the stars align it can be popped instead of the get.
Block removal - remove any blocks with no branches, as they are valid in wasm binary format.
Implementation-wise, the IR is defined in wasm-stack.h. A new StackInst is defined, representing a single instruction. Most are simple reflections of Binaryen IR (an add, a load, etc.), and just pointers to them. Control flow constructs are expanded into multiple instructions, like a block turns into a block begin and end, and we may also emit extra unreachables to handle the fact Binaryen IR has unreachable blocks/ifs/loops but wasm does not. Overall, all the Binaryen IR differences with wasm vanish on the way to stack IR.
Where this IR lives: Each Function now has a unique_ptr to stack IR, that is, a function may have stack IR alongside the main IR. If the stack IR is present, we write it out during binary writing; if not, we do the same binaryen IR => wasm binary process as before (this PR should not affect speed there). This design lets us use normal Passes on stack IR, in particular this PR defines 3 passes:
Generate stack IR
Optimize stack IR (might be worth splitting out into separate passes eventually)
Print stack IR for debugging purposes
Having these as normal passes is convenient as then they can run in parallel across functions and all the other conveniences of our current Pass system. However, a downside of keeping the second IR as an option on Functions, and using normal Passes to operate on it, means that we may get out of sync: if you generate stack IR, then modify binaryen IR, then the stack IR may no longer be valid (for example, maybe you removed locals or modified instructions in place etc.). To avoid that, Passes now define if they modify Binaryen IR or not; if they do, we throw away the stack IR.
Miscellaneous notes:
Just writing Stack IR, then writing to binary - no optimizations - is 20% slower than going directly to binary, which is one reason why we still support direct writing. This does lead to some "fun" C++ template code to make that convenient: there is a single StackWriter class, templated over the "mode", which is either Binaryen2Binary (direct writing), Binaryen2Stack, or Stack2Binary. This avoids a lot of boilerplate as the 3 modes share a lot of code in overlapping ways.
Stack IR does not support source maps / debug info. We just don't use that IR if debug info is present.
A tiny text format comment (if emitting non-minified text) indicates stack IR is present, if it is ((; has Stack IR ;)). This may help with debugging, just in case people forget. There is also a pass to print out the stack IR for debug purposes, as mentioned above.
The sieve binaryen.js test was actually not validating all along - these new opts broke it in a more noticeable manner. Fixed.
Added extra checks in pass-debug mode, to verify that if stack IR should have been thrown out, it was. This should help avoid any confusion with the IR being invalid.
Added a comment about the possible future of stack IR as the main IR, depending on optimization results, following some discussion earlier today.
|
| |
|
|
|
|
|
| |
On a codebase with 370K functions, 160K were in fact duplicate (!)... and it took many many passes to figure that out, over 2 minutes in fact (!), as A and B may be identical only after we see that the functions C1, C2 that they call are identical (so there can be long "chains" here).
To avoid this, limit how many passes we do. In -O1, just do one pass - that gets most duplicates. In -O2, do 10 passes - that gets almost all of it on this codebase. And in -O3 (or -Os/-Oz) do as many passes as necessary (i.e., the old behavior). This at least lets iteration builds (-O1) be nice and fast.
This PR also refactors the hashing code used in that pass, moving it to nicer header files for clearer readability. Also some other minor cleanups in hashing code that helped debug this.
|
|
|
The IR is indeed a tree, but not an "abstract syntax tree" since there is no language for which it is the syntax (except in the most trivial and meaningless sense).
|
