Skip to content

Latest commit

 

History

History
658 lines (555 loc) · 26.7 KB

CHANGELOG.md

File metadata and controls

658 lines (555 loc) · 26.7 KB
Raw

Changelog

This document describes changes between tagged Binaryen versions.

To browse or download snapshots of old tagged versions, visit https://github.com/WebAssembly/binaryen/releases.

Not all changes are documented here. In particular, new features, user-oriented fixes, options, command-line parameters, usage changes, deprecations, significant internal modifications and optimizations etc. generally deserve a mention. To examine the full set of changes between versions, visit the link to full changeset diff at the end of each section.

Current Trunk

v121

  • BinaryenSelect no longer takes a type parameter. (#7097)
  • AutoDrop APIs have been removed. (#7106)
  • bulk-memory-opt and call-indirect-overlong features are added for parity with LLVM. (#7139)
  • WasmGC optimizations now run significantly faster and scale better with available threads. (#7142)
  • Binaryen now supports parsing control flow structures with parameter types by lowering them away in the parsers. (#7149)

v120

  • Remove closed world validation checks. These checks were causing more harm than good. All valid code will now validate with --closed-world (but also it now provides fewer warnings to users that enable closed world on code which does not conform to the requirements of that mode, which can lead to changes in runtime behavior; for the long-term plans, see #6965). (#7019)
  • Many compile time speedups were implemented (2x overall improvement), see #4165 (comment)
  • Several exnref (newest version of Wasm EH) optimizations: #7013, #6996, #6997, #6983, #6980
  • Source Maps: Support 5 segment mappings. (#6795)
  • [wasm-split] Add a multi-split mode. (#6943)
  • Add a --preserve-type-order option that minimizes text format changes in type ordering. (#6916)
  • Add a J2CL specific pass that moves itable entries to vtables. (#6888)

v119

  • Passes can now receive individual pass arguments, that is --foo=A --foo=B for a pass foo will run the pass twice (which was possible before) and will now run it first with argument A and second with B. --pass-arg=foo@BAR will now apply to the most recent --foo pass on the commandline, if foo is a pass (while global pass arguments - that are not the name of a pass - remain, as before, global for all passes). (#6687)
  • The Metrics pass now takes an optional argument to use as the title, --metrics=text will show that text before the metrics. Each instance of Metrics can have unique text, --metrics=before -O3 --metrics=after. (#6792)
  • Add C and JS APIs to control more pass options (trapsNeverHappen, closedWorld, generateStackIR, optimizeStackIR, and the list of skipped passes). (#6713)
  • A C APIs for getting/setting the type of Functions (#6721).
  • Allow using --skip-pass on the commandline multiple times (#6714).
  • The instructions relaxed_fma and relaxed_fnma have been renamed to relaxed_madd and relaxed_nmadd.
  • Add a new --heap-store-optimization pass. (#6882)
  • Add a pass for minimizing recursion groups. (#6832)

v118

  • StackIR is now handled entirely during binary writing (#6568). This is mostly not noticeable, except that:
    • Text output no longer notes (; has Stack IR ;) (as Stack IR only exists during binary writing).
    • --generate-stack-ir, --optimize-stack-ir, and --print-stack-ir are now flags and not passes. That means the order of operations may seem different, as they apply during binary writing (or, if no binary is written but we were still asked to print StackIR, wasm-opt does it at the very end).
    • Whether to generate, optimize, and print StackIR is now noted as part of the PassOptions. As a result BinaryenModulePrintStackIR and similar APIs do not receive an optimize flag, as they read the PassOption optimizeStackIR instead.
  • The new, standards-compliant text parser is now the default.
  • Source map comments on else branches must now be placed above the instruction inside the else branch rather than on the else branch itself.
  • Source map locations from instructions are no longer automatically propagated to function epilogues.
  • Add a new BinaryenModuleReadWithFeatures function to the C API that allows to configure which features to enable in the parser. (#6380)
  • The build-time option to use legacy WasmGC opcodes is removed. (#5874)
  • The strings in string.const instructions must now be valid WTF-8.
  • The TraverseCalls flag for ExpressionRunner is removed.
  • C API: Support adding data segments individually (#6346)
  • Add sourcemap support to wasm-metadce and wasm-merge (#6372).
  • Fix semantics of return calls (#6448, #6451, #6464, #6470, #6474).
  • Add table64 lowering pass (#6595).
  • Add TraceCalls instrumentation pass (#6619).

v117

  • Add a WebAssembly build to release (#6351)
  • Add Linux aarch64 build to release (#6334).
  • The text format for tuple instructions now requires immediates. For example, to make a tuple of two elements, use tuple.make 2 (#6169) (#6172) (#6170).
  • The text format for if expressions now requires then and else to introduce the two branch arms, matching the spec (#6201).
  • Fuzzer: Remove --emit-js-shell logic and reuse fuzz_shell.js instead (#6310).
  • [EH] Add --experimental-new-eh option to wasm-opt (#6270) (#6210).
  • Add StringLowering pass, from stringref to imported-strings (#6271).
  • C API: Add BinaryenFunctionAppendVar (#6213).
  • Add J2CL optimization pass (#6151).
  • Add no-inline IR annotation, and passes to set it based on function name (#6146).
  • C API: Add BinaryenTableGetType and BinaryenTableSetType (#6137).
  • Add an Unsubtyping optimization (#5982).
  • Compute full transitive closure in GlobalEffects (#5992).
  • Add passes to finalize or unfinalize types (#5944).
  • Add a tuple optimization pass (#5937).

v116

  • "I31New" changed to "RefI31" everywhere it appears in the C API and similarly "i31.new" has been replaced with "ref.i31" in the JS API and in printed output (#5930, #3931).
  • The standard WasmGC opcodes are now on by default (#5873).

v115

  • The nonstandard, experimental gc-nn-locals feature has been removed now that standard non-nullable locals are supported. (#5080)
  • Print all debug annotations when BINARYEN_PRINT_FULL. (#5904)
  • Use the standard text format for WasmGC now that things are stable there. (#5894, #5889, #5852, #5801)
  • Rename multimemory flag to --multimemory to match its naming in LLVM. (#5890)
  • Allow opting into standard GC encodings at compile time. (#5868)
  • Remove legacy WasmGC instructions. (#5861)
  • GUFA: Infer using TrapsNeverHappen. (#5850)
  • Add a pass to reorder functions by name, which can be useful for debugging (e.g. comparisons after optimizations), --reorder-functions-by-name. (#5811)
  • C API: Add BinaryenAddFunctionWithHeapType which is like BinaryenAddFunction but takes a heap type. The old function is kept for backwards compatibility and as a convenience. (#5829)
  • Work on new optimization framework using lattices. (#5784, #5794, #5800, #5817, #5831, #5849)
  • Heap types are now final by default and openness must be opted into both in the text and binary formats as well as in the TypeBuilder API. (#5918)

v114

  • Fix a bug where e.g. -O3 -Oz ran the -O3 with the opt levels of -Oz, which could inhibit inlining, for example. While this is a bugfix, it affects how commandline options are interpreted, so if you depended on the old behavior this may be a breaking change. That is, the old behavior made -O3 -Oz run the first -O3 with -Oz's opt levels, and the new behavior is to run -O3 with the proper (-O3) opt levels. This is a followup to #5333 from a previous release. (#5787)
  • Add pass to remove Exception Handling instructions (#5770).

v113

  • Add a wasm-merge tool. This is a full rewrite of the previous wasm-merge tool that was removed from the tree in the past. The new version is much simpler after recent improvements to multi-memory and multi-table. The rewrite was motivated by new use cases for merging modules in the context of WasmGC.
  • Some C and JS API functions now refer to data and element segments by name instead of index.
  • The --nominal and --hybrid command line options and related API functions have been removed. The only supported type system is now the standard isorecursive (i.e. hybrid) type system. (#5672)
  • Add a "mayNotReturn" effect (#5711).
  • Disable the memory64 feature in Memory64Lowering.cpp (#5679).
  • Disable sign extension in SignExtLowering.cpp (#5676).

v112

  • Add AbstractTypeRefining pass (#5461)
  • Add a mechanism to skip a pass by name (#5448)
  • Add TypeMerging pass (#5321)
  • Add TypeSSA pass (#5299)
  • Optimization sequences like -O3 -Os now do the expected thing and run -O3 followed by -Os. Previously the last of them set the defaults that were used by all executions, so -O3 -Os was equivalent to -Os -Os. (There is no change to the default optimization level that other passes can see. For example, --precompute-propagate -O2 -O1 will run --precompute-propagate at opt level 1, as the global default is set to 2 and then overridden to 1. The only change is that the passes run by -O2 will actually run -O2 now, while before they'd use the global default which made them do -O1.)
  • Add --closed-world flag. This enables more optimizations in GC mode as it lets us assume that we can change types inside the module.
  • The isorecursive WasmGC type system (i.e. --hybrid) is now the default to match the spec and the old default equirecursive (i.e. --structural) system has been removed.
  • ref.is_func, ref.is_data, and ref.is_i31 have been removed from the C and JS APIs and RefIs has been replaced with RefIsNull.
  • Types Data and Dataref have been replaced with types Struct and Structref in the C and JS APIs.
  • BinaryenStringNew now takes an additional last argument, try_, indicating whether the instruction is one of string.new_utf8_try respectively string.new_utf8_array_try.
  • BinaryenStringEq now takes an additional second argument, op, that is either BinaryenStringEqEqual() if the instruction is string.eq or BinaryenStringEqCompare() if the instruction is string.compare.

v111

  • Add extra memory64 argument for BinaryenSetMemory and new BinaryenMemoryIs64 C-API method to determine 64-bit memory. (#4963)
  • TypeBuilderSetSubType now takes a supertype as the second argument.
  • call_ref now takes a mandatory signature type immediate.
  • If THROW_ON_FATAL is defined at compile-time, then fatal errors will throw a std::runtime_error instead of terminating the process. This may be used by embedders of Binaryen to recover from errors.
  • Implemented bottom heap types: none, nofunc, and noextern. RefNull expressions and null Literals must now have type nullref, nullfuncref, or nullexternref.
  • The C-API's BinaryenTypeI31ref and BinaryenTypeDataref now return nullable types.
  • The sign-extension and mutable-globals features are now both enabled by default in all tools. This is in order to match llvm's defaults (See https://reviews.llvm.org/D125728).
  • Add a pass to lower sign-extension operations to MVP.

v110

  • Add support for non-nullable locals in wasm GC. (#4959)
  • Add support for multiple memories. (#4811)
  • Add support for the wasm Strings proposal. (see PRs with [Strings] in name)
  • Add a new flag to Directize, --pass-arg=directize-initial-contents-immutable which indicates the initial table contents are immutable. That is the case for LLVM, for example, and it allows us to optimize more indirect calls to direct ones. (#4942)
  • Change constant values of some reference types in the C and JS APIs. This is only observable if you hardcode specific values instead of calling the relevant methods (like BinaryenTypeDataref()). (#4755)
  • BinaryenModulePrintStackIR, BinaryenModuleWriteStackIR and BinaryenModuleAllocateAndWriteStackIR now have an extra boolean argument optimize. (#4832)
  • Remove support for the let instruction that has been removed from the typed function references spec.
  • HeapType::ext has been restored but is no longer a subtype of HeapType::any to match the latest updates in the GC spec. (#4898)
  • i31ref and dataref are now nullable to match the latest GC spec. (#4843)
  • Add support for extern.externalize and extern.internalize. (#4975)

v109

  • Add Global Struct Inference pass (#4659) (#4714)
  • Restore and fix SpillPointers pass (#4570)
  • Update relaxed SIMD instructions to latest spec

v108

  • Add CMake flag BUILD_TOOLS to control building tools (#4655)
  • Add CMake flag JS_OF_OCAML for js_of_ocaml (#4637)
  • Remove externref (#4633)

v107

  • Update the wasm GC type section binary format (#4625, #4631)
  • Lift the restriction in liveness-traversal.h on max 65535 locals (#4567)
  • Switch to nominal fuzzing by default (#4610)
  • Refactor Feature::All to match FeatureSet.setAll() (#4557)
  • New Signature Pruning pass (#4545)
  • Add support for extended-const proposal (#4529)
  • Add BUILD_TESTS CMake option to make gtest dependency optional.
  • Updated tests to use filecheck 0.0.22 (#4537). Updating is required to successfully run the lit tests. This can be done with pip3 install -r requirements-dev.txt.

v106

  • [wasm2js] Support exports of Globals (#4523)
  • MergeSimilarFunctions optimization pass (#4414)
  • Various wasm-ctor-eval improvements, including support for GC.

v105

  • This release contains binaries for ARM64 MacOS devices (#4397)
  • Otherwise, mostly bug fixes and incremental optimization improvements.

v104

  • Bugfixes only, release created due to incorrect github release artifacts in v103 release (#4398).

v103

  • The EffectAnalyzer now takes advantage of immutability of globals. To achieve that it must have access to the module. That is already the case in the C++ API, but the JS API allowed one to optionally not add a module when calling getSideEffects(). It is now mandatory to pass in the module.
  • JS and Wasm builds now emit ECMAScript modules. New usage is: 
    import Binaryen from "path/to/binaryen.js";
const binaryen = await Binaryen();
...
  • CallIndirect changed from storing a Signature to storing a HeapType

v102

  • Add BinaryenUpdateMaps to the C API.

  • Adds a TrapsNeverHappen mode (#4059). This has many of the benefits of IgnoreImplicitTraps, but can be used safely in more cases. IgnoreImplicitTraps is now deprecated.

  • Adds type argument for BinaryenAddTable method (#4107). For the binaryen.js api this parameter is optional and by default is set to funcref type.

  • Replace BinaryenExpressionGetSideEffects's features parameter with a module parameter.

  • OptimizeInstructions now lifts identical code in select/if arms (#3828). This may cause direct BinaryenTupleExtract(BinaryenTupleMake(...)) to use multivalue types.

v101

  • BinaryenSetFunctionTable and module.setFunctionTable have been removed in favor of BinaryenAddTable and module.addTable respectively.
  • BinaryenIsFunctionTableImported is removed.
  • A new type BinaryenElementSegmentRef has been added to the C API with new apis in both C & JS:
    • BinaryenAddActiveElementSegment
    • BinaryenAddPassiveElementSegment
    • BinaryenRemoveElementSegment
    • BinaryenGetElementSegment
    • BinaryenGetElementSegmentByIndex
    • BinaryenElementSegmentGetName
    • BinaryenElementSegmentSetName
    • BinaryenElementSegmentGetTable
    • BinaryenElementSegmentSetTable
    • BinayenElementSegmentIsPassive
    • module.addActiveElementSegment
    • module.addPassiveElementSegment
    • module.removeElementSegment
    • module.getElementSegment
    • module.getElementSegmentByIndex
    • module.getTableSegments
    • module.getNumElementSegments
    • binaryen.getElementSegmentInfo
  • BinaryenAddTable and module.addTable no longer take offset and function names.
  • BinaryenGetNumFunctionTableSegments is replaced with BinaryenGetNumElementSegments.
  • BinaryenGetFunctionTableSegmentOffset is replaced with BinaryenElementSegmentGetOffset.
  • BinaryenGetFunctionTableSegmentLength is replaced with BinaryenElementSegmentGetLength.
  • BinaryenGetFunctionTableSegmentData is replaced with BinaryenElementSegmentGetData.
  • Boolean values in the C API now should use bool instead of int.
  • Experimental SIMD instructions have been removed and the names and opcodes of the standard instructions have been updated to match the final spec.

v100

  • wasm-dis now supports options to enable or disable Wasm features.
  • Reference types support has been improved by allowing multiple tables in a module.
  • call_indirect and return_call_indirect now take an additional table name parameter. This is necessary for reference types support.
  • New getter/setter methods have been introduced for call_indirect table name:
    • BinaryenCallIndirectGetTable
    • BinaryenCallIndirectSetTable
    • JS API CallIndirect.table
  • New APIs have been added to add and manipulate multiple tables in a module:
    • BinaryenAddTable
    • BinaryenRemoveTable
    • BinaryenGetNumTables
    • BinaryenGetTable
    • BinaryenGetTableByIndex
    • BinaryenTableGetName
    • BinaryenTableGetInitial
    • BinaryenTableHasMax
    • BinaryenTableGetMax
    • BinaryenTableImportGetModule
    • BinaryenTableImportGetBase
    • module.addTable
    • module.removeTable
    • module.getTable
    • module.getTableByIndex
    • module.getNumTables
    • binaryen.getTableInfo

v99

  • Fix optimization behavior on assuming memory is zero-filled. We made that assumption before, but it is incorrect in general, which caused problems. The fixed behavior is to not assume it, but require the user to pass it in as a flag, --zero-filled-memory. Large binaries with lots of empty bytes in the data section may regress without that flag. Toolchains like Emscripten can pass the flag automatically for users if they know it is right to assume, which can avoid any regressions. (#3306)
  • RefFunc C and JS API constructors (BinaryenRefFunc and ref.func respectively) now take an extra type parameter, similar to RefNull. This is necessary for typed function references support.
  • JS API functions for atomic notify/wait instructions are renamed.
    • module.atomic.notify -> module.memory.atomic.notify
    • module.i32.atomic.wait -> module.memory.atomic.wait32
    • module.i64.atomic.wait -> module.memory.atomic.wait64
  • Remove old/broken SpollPointers pass. This pass: Spills values that might be pointers to the C stack. This allows Boehm-style GC to see them properly. This can be revived if needed from git history (#3261).
  • Make NUM_PARAMS in FuncCastEmulation a runtime configuration option named max-func-params. This defaults to the original value of 16.
  • BinaryenGetFunction, BinaryenGetGlobal and BinaryenGetEvent now return NULL instead of aborting when the respective element does not yet exist.

v98

  • Add --fast-math mode. (#3155)
  • Initial implementation of "Memory64" proposal (#3130)
  • Lots of changes in support of GC proposal

v97

  • Remove asm2wasm, which supported Emscripten's fastcomp backend, after fastcomp was removed.
  • The new feature flag --enable-anyref enables just the anyref type incl. basic subtyping of externref, funcref and exnref (if enabled).
  • Enabling the exception handling or anyref features without also enabling reference types is a validation error now.
  • The Host expression and its respective APIs have been refactored into separate MemorySize and MemoryGrow expressions to align with other memory instructions.

v96

  • Fuzzing: Compare wasm2js to the interpreter (#3026)
  • Fix CountLeadingZeroes on MSVC, which lead to bad optimizations (#3028)
  • Asyncify verbose option (#3022)
  • wasm2js: Add an "Export" scope for name resolution, avoids annoying warnings (#2998)
  • Extend the C- and JS-APIs (#2586)

v95

  • Add Asyncify "add list" that adds to the list of functions to be instrumented. Rename old lists to be clearer and more consistent with that, so now there is "remove list" to remove, "add list" to add, and "only list" which if set means that only those functions should be instrumented and nothing else.
  • Renamed various ambiguous C-API functions for consistency:
    • BinaryenBlockGetChild to BinaryenBlockGetChildAt
    • BinaryenSwitchGetName to BinaryenSwitchGetNameAt
    • BinaryenCallGetOperand to BinaryenCallGetOperandAt
    • BinaryenCallIndirectGetOperand to BinaryenCallIndirectGetOperandAt
    • BinaryenHostGetOperand to BinaryenHostGetOperandAt
    • BinaryenThrowGetOperand to BinaryenThrowGetOperandAt
    • BinaryenTupleMakeGetOperand to BinaryenTupleMakeGetOperandAt

v94

  • The C-API's BinaryenSetAPITracing and the JS-API's setAPITracing have been removed because this feature was not very useful anymore and had a significant maintainance cost.
  • wasm-emscripten-finalize will no longer generate stackSave, stackAlloc, stackRestore function. It not expects them to be included in the input file.

v93

  • First release with binaries built with github actions.

v92

  • The multivalue feature has been added. It allows functions and control flow structures to return tuples and for locals and globals to have tuple types. Tuples are created with the new tuple.make pseudoinstruction and their elements are retrieved with the new tuple.extract pseudoinstruction.
  • The internal type interner has been rewritten to avoid taking locks in far more situations. Depending on the workload, this may result in large speedups and increased parallelism.
  • Represent the dylink section in Binaryen IR, so we can read, write, and update it.

v91

  • BinaryenExpressionGetSideEffects (C API) and getSideEffects (JS API) now takes an additional features parameter.
  • Reference type support is added. Supported instructions are ref.null, ref.is_null, ref.func, and typed select. Table instructions are not supported yet. For typed select, C/JS API can take an additional 'type' parameter.

v90

  • local.tee's C/Binaryen.js API now takes an additional type parameter for its local type, like local.get. This is required to handle subtypes.
  • Added load_splat SIMD instructions
  • Binaryen.js instruction API changes:
    • notify -> atomic.notify
    • i32.wait / i64.wait -> i32.atomic.wait / i64.atomic.wait
  • Binaryen.js: flags argument in setMemory function is removed.
  • atomic.fence instruction support is added.
  • wasm-emscripten-finalize: Don't rely on name section being present in the input. Use the exported names for things instead.
  • Added mutable parameter to BinaryenAddGlobalImport.
  • Replace BinaryenSIMDBitselect* with BinaryenSIMDTernary* in the C API and add qfma/qfms instructions.
  • Added offset parameter to BinaryenSetFunctionTable.
  • Add the ability to create multivalue Types in the C and JS APIs.
  • Remove named function types. They are replaced by params and results types local to each function.
  • Binaryen.js can now be compiled to Wasm using the binaryen_wasm target. Unlike the JS variant, the Wasm variant requires asynchronously awaiting the Wasm blob's instantiation and initialization before being usable, using the binaryen.ready promise, e.g. binaryen.ready.then(() => ...).
  • Binaryen.js now uses binaryen (was Binaryen) as its global name to align with the npm package.
  • Binaryen.js: The result of getMemorySegmentInfoByIndex now has the same structure as the respective inputs on creation (byteOffset -> offset).

v88

  • wasm-emscripten-finalize: For -pie binaries that import a mutable stack pointer we internalize this an import it as immutable.
  • The tail-call feature including the return_call and return_call_indirect instructions is ready to use.

v87

  • Rename Bysyncify => Asyncify

v86

  • The --initial-stack-pointer argument to wasm-emscripten-finalize no longer has any effect. It will be removed completely in future release.

v85

  • Wast file parsing rules now don't allow a few invalid formats for typeuses that were previously allowed. Typeuse entries should follow this format, meaning they should have (type) -> (param) -> (result) order if more than one of them exist. 
    typeuse ::= (type index|name)+ |
            (type index|name)+ (param ..)* (result ..)* |
            (param ..)* (result ..)*
    Also, all (local) nodes in function definition should be after all typeuse elements.
  • Removed APIs related to deprecated instruction names in Binaryen.js:
    • get_local / getLocal
    • set_local / setLocal
    • tee_local / teeLocal
    • get_global / getGlobal
    • set_global / setGlobal
    • current_memory / currentMemory
    • grow_memory / growMemory They are now available as their new instruction names: local.get, local.set, local.tee, global.get, global.set, memory.size, and memory.grow.
  • Add feature handling to the C/JS API with no feature enabled by default.

v84

  • Generate dynCall thunks for any signatures used in "invoke" calls.

v81

  • Fix AsmConstWalker handling of string address in arg0 with -fPIC code

v80

  • Change default feature set in the absence of a target features section from all features to MVP.

v79

  • Improve support for side modules

v78

  • Add namedGlobals to metadata output of wasm-emscripten-finalize
  • Add support for llvm PIC code.
  • Add --side-module option to wasm-emscripten-finalize.
  • Add segmentPassive argument to BinaryenSetMemory for marking segments passive.
  • Make -o - print to stdout instead of a file named "-".

v73

  • Remove wasm-merge tool.

v73

v55

  • RelooperCreate in the C API now has a Module parameter, and RelooperRenderAndDispose does not.
    • The JS API now has the Relooper constructor receive the Module.
  • Relooper: Condition properties on Branches must not have side effects.

older

  • BinaryenSetFunctionTable in the C API no longer accepts an array of functions, instead it accepts an array of function names, const char** funcNames. Previously, you could not include imported functions because they are of type BinaryenImportRef instead of BinaryenFunctionRef. #1650

  • BinaryenSetFunctionTable in the C API now expects the initial and maximum table size as additional parameters, like BinaryenSetMemory does for pages, so tables can be grown dynamically. #1687

  • Add shared parameters to BinaryenAddMemoryImport and BinaryenSetMemory, to support a shared memory. #1686