| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
per-memory (#2507)
The PR updates the bulk memory operations (memory.fill, memory.copy,
table.fill, etc.) to support 64-bit addresses and counts. Previously these functions
only took u32's, even with memory64 enabled. (#2506)
This PR also allows "software-bounds-checked" memories and "guard-page-checked"
memories to coexist in the same module. It creates two versions of every memory
operation: an unrestricted version (that works with any memory) and a _default32
version (for memories with default page size and i32 indexing). (#2507)
#2506 and #2507 have been squashed together to avoid a performance regression.
This is a stepping stone to supporting custom-page-sizes (which will need to be
software-bounds-checked) (#2508).
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
| |
This continues the work from #1783 and reduces special handling of elem
exprs, by treating them the same as other const expressions (init
expressions).
|
| | |
|
| | |
|
| |
|
| |
Co-authored-by: Shravan Narayan <shravanrn@gmail.com>
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* wasm2c: multiple .c outputs
This enables wasm2c to have multiple .c outputs, which allows parallel
compilation of the wasm2c outputs. This is useful when the input WASM module is
big.
wasm2c takes the number of .c outputs as an argument to `--num-outputs`
(defaulting to 1). If the number is equal to 1, the .c output does not change
except for two new macro declarations and the ordering of declarations and
definitions. If greater than 1, wasm2c outputs change in the following ways:
1) wasm2c outputs a [module-name]-impl.h that includes any module-wide
declarations, including:
* content of `WriteSourceTop()`
* function type declarations
* tag types
* tag declarations
* function declarations
* data segments and elem segments declarations
Any static declaration become extern in this header.
2) wasm2c outputs [module-name]_i.c for i = [0, ..., number of .c outputs - 1). Any
module-wide material is written to [module-name]_0.c, including:
* function types, tags, data segments, elem segments
* imports and exports
* module initialization, instantiation and free
3) For each function implementation, wasm2c assigns it to one output .c file
by sorting the function names and partitioning into roughly equal buckets.
Alternately, the caller can supply its own assignment function (helpful if it wants
the assignments to be more stable in the face of function insertion or deletion).
|
| |
|
|
|
|
| |
Add kLocalSymbolPrefix which is used for names of params, locals and
stack vars. This allows c-writer to not assign global_sym_map_ to
local_sym_map_ for writing each individual function, since local names
can't duplicate global names.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* wasm2c: prettify/change name-mangling
This refactors the wasm2c name-mangling in two big ways:
1) Removing the `Z_` prefix and trying to make the names somewhat
ergonomic/pretty. Previously the `factorial` export from a `fac`
module looked like this:
```
u32 Z_facZ_factorial(Z_fac_instance_t*, u32);
```
After this commit, it looks like this:
```
u32 w2c_fac_factorial(w2c_fac*, u32);
```
Symbols defined by wasm2c itself (including instantiate, free,
get_func_type and the imported memory limits) are now prefixed with
`wasm2c_` to avoid conflicting with names defined by the module.
2) Using globally unique (module-prefixed) names for functions, types,
segments, and tags, even though they are currently static
(internal-linkage) symbols in the .c output. This is preparation for
a future "multiple .c output" option where these symbols will need to
have external linkage.
|
| |
|
| |
This change incorporates [simd-everywhere](https://github.com/simd-everywhere/simde) into the wasm2c output, which maps wasm SIMD C intrinsics to any supported target architecture.
|
| |
|
|
|
|
| |
We had been running the GitHub UBSAN wasm2c tests with -fsanitize=undefined
but without -fno-sanitize-recover, meaning some of the spec tests
were printing UBSAN error messages but still returning 0, so we
weren't seeing the test failures.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This makes wasm2c serialize each function type, rather than registering
function types at module-initialization time. The serialized function
type is the SHA-256 of the mangled param and result types (with a space
between params and results).
At runtime in call_indirect, a known (immediate) function type is
compared against the function type stored in a funcref structure. For
call_indirects to functions local to the module, or for any
call_indirect when the toolchain merges string constants across
compilation units (generally, GCC and clang), this can be done by
comparing the pointers to each function type. Otherwise, the actual
32-byte values are compared.
The function type IDs can be looked up at runtime with
`Z_[modname]_get_func_type`, which matches the API from
`wasm_rt_register_func_type`. A new `callback` example demos this.
wasm2c does the SHA-256 either by linking against libcrypto or, if not
available or if requested via `cmake -DUSE_INTERNAL_SHA256=ON`, by using
a vendored (header-only) PicoSHA2. There is no runtime dependency on
SHA-256 in the wasm2c runtime or generated modules.
This eliminates the last of the per-module state, so this commit also removes
the [modname]_init_module() function and the s_module_initialized bool.
|
| |
|
|
| |
(#2133)
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As well as the testsuite update there are two notable changes that
come with it here. These can both be split out an landed first if
it makes sense.
1. wasm2c now supports element sections containing externref. Currently
only the null reference is supported.
2. element segments no longer use funcref as the default element type
but instead, unless explicitly included in the binary, the
element type defaults to the type of the table in which the segment
is active.
Fixes: #1612 #2022
|
| |
|
|
| |
Restores current versions of all non-SIMD tests in the core testsuite
and multi-memory and exception-handling proposals.
|
| |
|
|
| |
also: - cleanup handling of newlines
- "init_memory"/"init_table" -> "init_memories"/"init_tables"
|
| |
|
| |
Co-authored-by: Yuhan Deng <yhdeng@stanford.edu>
|
| |
|
| |
Co-authored-by: Angela Montemayor <amontema@cs.stanford.edu>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This change avoids some confusing/redundant bracketing and makes
the output a little easier to easy and more like regular C usage.
Old:
```
wasm_rt_memory_t (*Z_test_hello_worldZ_memory);
...
Z_test_hello_worldZ_memory = (&w2c_memory);
Z_test_hello_worldZ__start = (&w2c__start)
```
New:
```
wasm_rt_memory_t* Z_test_hello_worldZ_memory;
...
Z_test_hello_worldZ_memory = &w2c_memory;
Z_test_hello_worldZ__start = &w2c__start;
```
|
| |
|
|
|
| |
(The imports.txt and exports.txt tests require multi-table and
are disabled until reference types is supported.)
|
| |
|
|
|
|
|
|
|
| |
Enable optimization when compiling the wasm2c output on non-Windows platforms (effectively GCC and clang). This required:
- Preventing load instructions from being optimized away if their value is unused (using inline assembly with an input operand and empty code). This is necessary to force an OOB trap on platforms that use mprotect and the signal handler to detect OOB.
- Disabling tail-call optimization in the compiler, to make sure that infinite recursion traps. (This required bumping the version of macOS in GitHub Actions to get a new-enough AppleClang. We should revert this back to 'macos-latest' as soon as that becomes the default.)
- Using NaN-quieting versions of a bunch of FP ops that were previously only used on Windows, and adding floor/ceil and promotion/demotion.
- Using the '-frounding-math' and '-fsignaling-nans' compiler flags to tell GCC and clang not to fold certain FP ops (e.g. subtracting zero, multiplying by 1).
Fixes #1925.
|
| | |
|
| |
|
|
|
|
| |
This will be required by module instancing and bulk memory
ops to avoid a memory leak detected by asan.
Update README.md to reflect changes, including the `fac` walkthrough.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The module prefix is no longer optional and is now set by default to the
name of the module as given in the name section, or taken from the name
of the input file.
A new `-n/--module-name` command line flag can also be used to override.
The prefix used by the generated code is now fixed at wasm2c time which
makes the output code easier to read and also avoid the symbol pasting
in the C pre-processor which makes the source hard to understand. For
example, it makes symbols hard to `grep` for.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
| |
This effectively means that we no longer support imports that are
overloaded by signature only. This is not something that we need to
support in order to support the core wasm spec.
This feature is available in the JS embedding but there is no good
reason (AFAICT) to support it in wasm2c, and this simplifies the
generated code.
Fixes #1858
|
| |
|
|
|
|
| |
On MacOS, merge OOB and exhaustion traps. (Linux distinguishes these as SEGV_ACCERR vs. SEGV_MAPERR in the
si_code, but MacOS seems to deliver a SEGV_ACCERR for both.)
Add wasm_rt_init() and wasm_rt_free() functions to wasm-rt.h that the embedder must call to set up and clean up the runtime state. (If the embedder doesn't call these, OOB and exhaustion will result in an uncaught segfault.)
|
| |
|
|
|
|
|
| |
All tests are now passing with cl.exe under x64. With x86 there are some test failure that I believe relate
the use of the x87 registers to pass floating point numbers. I suggest we look into fixing those as a followup.
Split out from #1833
|
| |
|
|
|
| |
Do we really want to have this output checked in like this?
If so, we should probably automate its update ot at least add
a check to avoid it getting out-of-date.
|
| |
|
|
|
|
|
|
|
|
| |
This example demonstrates how to use imported functions. The `rot13`
program takes each command line argument, and rot13-encodes it.
The exported `rot13` function has no arguments, and instead calls back
into the program (via `fill_buf`) with a buffer to fill in. When the
function finishes it calls `buf_done`.
(rot13.wat is the same as in src/test-interp.cc.)
|
|
|
* Described how to use it, and put it near the top of the doc
* Added the sources and generated files
|
