currently this is not necessary because context (WASINNContext) is
local to instance. (wasm_module_instance_t)
i plan to make a context shared among instances in a cluster when
fixing https://github.com/bytecodealliance/wasm-micro-runtime/issues/4313.
this is a preparation for that direction.
an obvious alternative is to tweak the module instance context APIs
to allow declaring some kind of contexts instance-local. but i feel,
in this particular case, it's more natural to make "wasi-nn handles"
shared among threads within a "process".
note that, spec-wise, how wasi-nn behaves wrt threads is not defined
at all because wasi officially doesn't have threads yet. i suppose, at
this point, that how wasi-nn interacts with wasi-threads is something
we need to define by ourselves, especially when we are using an outdated
wasi-nn version.
with this change, if a thread attempts to access a context while
another thread is using it, we simply make the operation fail with
the "busy" error. this is intended for the mimimum serialization to
avoid problems like crashes/leaks/etc. this is not intended to allow
parallelism or such.
no functional changes are intended at this point yet.
cf.
https://github.com/bytecodealliance/wasm-micro-runtime/issues/4313https://github.com/bytecodealliance/wasm-micro-runtime/issues/2430
copied from the linux version.
i'm a bit skeptical with this workaround though.
it might be simpler to prohibit the use of wamr api in these
shared libraries. after all, what these libraries do is nothing
specific to wasm.
an example application with flexible cli options which
aims to allow us to perform any wasi-nn operations.
eg.
```
--load-graph=file=fixture/model.xml,file=fixture/model.bin,id=graph
--init-execution-context=graph-id=graph,id=ctx
--set-input=file=fixture/tensor.bgr,context-id=ctx,dim=1,dim=3,dim=224,dim=224
--compute=context-id=ctx
--get-output=context-id=ctx,file=output.bin
```
as wasi-nn doesn't have these concepts, the best we can do without
risking breaking certain applications here is to pass through tensors
as they are.
this matches wasmtime's behavior.
tested with:
* wasmtime classification-example
(with this change, this example fails on tensor size mismatch
instead of implicitly resizing it.)
* license-plate-recognition-barrier-0007, a converted version
with non-fp32 output. [1]
(with this change, this model outputs integers as expected.)
[1] https://github.com/openvinotoolkit/open_model_zoo/tree/cd7ebe313b69372763e76b82e5d24935308fece4/models/public/license-plate-recognition-barrier-0007
- wasi_nn.h: make this compatible with wasi_ephemeral_nn
cf. https://github.com/bytecodealliance/wasm-micro-runtime/issues/4323
- fix WASM_ENABLE_WASI_EPHEMERAL_NN build
this structure is used by host logic as well.
ideally definitions for wasm and host should be separated.
until it happens, check __wasm__ to avoid the breakage.
* this enum is (@witx tag u8) in witx
* it seems that some wasm modules actually use non-zero padding
and cause errors
* it's a bad practice to use C enum for ABI description anyway
sync with a more appropriate version of the definitions.
as we use the "wasi_ephemeral_nn", which is p1-based, it seems
more appropriate to use definitions from witx, not wit.
it's a bit unfortunate p2-based wasi-nn made gratuitous changes
like this from p1.
note: this is an ABI change.
it seems meaningless and quite confusing to access a table with
two aliases ("lookup" and "backends") within a function.
no functional changes are intended.
This partly reverts "Cmake improvements".
(https://github.com/bytecodealliance/wasm-micro-runtime/pull/4076)
Recently we changed the location to install public headers.
For example,
Old: include/wasm_export.h
New: include/iwasm/wasm_export.h
For cmake-based user applications using find_package(iwasm),
the cmake package, namely target_include_directories(INSTALL_INTERFACE),
is expected to add necessary compiler options like -isystem automatically.
(See samples/printversion for an example of such user applications.)
However, in reality, not every user application uses cmake.
This commit reverts the location to install public headers
for now, to avoid breakage for non-cmake user applications.
In case we want to re-apply the location change in future,
we should better communicate to the users. (eg. document
migration proceduces in release notes.)
Fixes: https://github.com/bytecodealliance/wasm-micro-runtime/issues/4290
References:
https://cmake.org/cmake/help/latest/prop_tgt/INTERFACE_INCLUDE_DIRECTORIES.html
As far as I know, we don't implement the proposal at all.
```
spacetanuki% wasm2wat --enable-all data.28.wasm
(module
(memory (;0;) 1)
(data (;0;) (i32.const 42
i32.const 0
i32.sub) ""))
spacetanuki% toywasm --load data.28.wasm
spacetanuki% ~/git/wasm-micro-runtime/product-mini/platforms/darwin/b.classic/iwasm data.28.wasm
WASM module load failed: illegal opcode or constant expression required or type mismatch
spacetanuki%
```
data.28.wasm in the above example is a binary version of:
https://github.com/WebAssembly/extended-const/blob/8d4f6aa2b00a8e7c0174410028625c6a176db8a1/test/core/data.wast#L184-L187
- it's unused
- valgrind is basically a linux-only software.
it isn't a good idea to make it a hard requirement.
if we want to use valgrind, it's better to introduce
a separate option to control it.
LLVM 16 and later expands cttz intrinsic to a table lookup,
which involves some relocations. (unless ZBB is available,
in which case the native instructions are preferred over
the table-based lowering.)
cf. https://reviews.llvm.org/D128911
the corresponding LLVM intrinsics' return types are same as
their first argument. eg. i64 for llvm.cttz.i64.
cf. https://llvm.org/docs/LangRef.html#llvm-cttz-intrinsic
this commit changes the return types of our versions of the
intrinsics to match llvm versions as our aot compiler,
specifically __call_llvm_intrinsic, assumes.
strictly speaking, this is a potential AOT ABI change.
however, I suppose it isn't a problem for many of 64-bit ABIs
out there, where (lower half of) a 64-bit register is used to
return a 32-bit value anyway. (for such ABIs, this commit
would fix the upper 32-bit value of the register.)
LLVM 19 and later started to use srodata ("small read only data")
sections for RISCV. cf. https://github.com/llvm/llvm-project/pull/82214
this commit makes our aot emitter/loader deal with those sections.
an alternative would be to disable small data sections completely by
setting the "SmallDataLimit" module attribute to zero. however, i feel
this commit is more straightforward and consisitent as we are already
dealing with sdata sections.
while ideally a user should not need to care this kind of
optimization details, in reality i guess it's sometimes useful.
both of clang and GCC expose a similar option. (-fno-jump-tables)
LLVM 18 and later, instcombine perfoms only one iteration.
it performs extra "verify fixpoint" operation when instcombine
is specified in certain ways, including how we do so here.
a problem is that the verification raises a fatal error when it
finds we didn't reach a fixpoint:
LLVM ERROR: Instruction Combining did not reach a fixpoint
after 1 iterations
while it should be rare, it's quite normal not to reach a fixpoint.
this commit fixes the issue by simply disabing the verification.
cf. https://github.com/llvm/llvm-project/commit/41895843b5915bb78e9d02aa711fa10f7174db43
In call_wasm_with_hw_bound_check/call_native_with_hw_bound_check,
ensure to set up the stack boundary (wasm_exec_env_set_thread_info)
before checking the overflow.
It seems that the problem was introduced by:
https://github.com/bytecodealliance/wasm-micro-runtime/pull/2940
This fixes link errors seen on my environment.
(macOS 15.2, x86-64, Xcode 16.2)
Tested as:
```
mkdir b
cd b
cmake -D WAMR_BUILD_JIT=1 -D LLVM_DIR=/usr/local/opt/llvm@19/lib/cmake/llvm ..
make
```
* Bump AOT_CURRENT_VERSION for WAMR 2.x (gc, memory64)
Maybe it's too late because we have already made a few releases
since then.
But this might still help users who haven't upgraded to WAMR 2.x yet.
Also, for the purpose of the versioning, it's safer to bump
needlessly than missing necessary bumps.
Fixes https://github.com/bytecodealliance/wasm-micro-runtime/issues/3837
* test-tools/aot-analyzer/include/config.h: bump AOT_CURRENT_VERSION
I'm not sure we want to use C99 %tu here.
While C99 %zu is more widely used in WAMR, %tu is rare (if any)
and I'm not sure if it's ubiquitously implemented in platforms
we support.
For JIT, we naturally use mach-o on macOS, where the section name
we currently use is not valid and ends up with the errors like:
```
LLVM ERROR: Global variable '__orc_lcl.aot_stack_sizes.0' has an invalid section specifier '.aot_stack_sizes': mach-o section specifier requires a segment and section separated by a comma.
```
Because the dedicated section is not necessary for JIT,
this commit simply stops using it.
Fixes: https://github.com/bytecodealliance/wasm-micro-runtime/issues/3730
This allows users to specify llvm command line options, similarly to clang's -mllvm option.
My motivations:
* -debug and friends
* -mtext-section-literals for xtensa
The current coding assumes the tmp file is created in /tmp.
However,
* It's somewhere under /var/folders/... by default on macOS.
* A user can set TMPDIR.
The "handlers" on the interpreter stack is sometimes treated as
host pointers and sometimes treated as i64 values. It's quite broken
for targets where pointers are not 64-bit.
This commit makes them host pointers consistently. (at least for
32-bit and 64-bit pointers. We don't support other pointer
sizes anyway.)
Fixes https://github.com/bytecodealliance/wasm-micro-runtime/issues/3110
aot_load_const_from_table() hides the const-ness of the
value and prevents optimizations like
https://github.com/bytecodealliance/wasm-micro-runtime/pull/3552.
This commit makes the aot compiler tracks the const-ness
of the value directly in the AOTValue and enables the above
mentioned optimization for XIP.
* I believe that LLVM MemoryBuffer interface is supposed to be read-only
and it's allowed to use eg. read-only mmap of the underlying file.
It isn't appropriate to modify the view at all.
* in case of WASM_ENABLE_DEBUG_AOT, the whole buffer is written as the text
section of the aot file. the modified e_type would confuse dwarf consumers.
note that, even when we are using XIP, the debug info usually contains
relocations. for example, llvm-dwarfdump doesn't seem to perform relocations
on .debug_info section for ET_CORE (== 4 == our E_TYPE_XIP) objects.
Consider the following wasm module:
```wast
(module
(func (export "foo")
i32.const 0x104
i32.const 0x12345678
i32.store
)
(memory 1 1)
)
```
While the address (0x104) is perfectly aligned for i32.store,
as our aot compiler uses 1-byte alignment for load/store LLVM
IR instructions, it often produces inefficient machine code,
especially for alignment-sensitive targets.
For example, the above "foo" function is compiled into the
following xtensa machine code.
```
0000002c <aot_func_internal#0>:
2c: 004136 entry a1, 32
2f: 07a182 movi a8, 0x107
32: 828a add.n a8, a2, a8
34: 291c movi.n a9, 18
36: 004892 s8i a9, a8, 0
39: 06a182 movi a8, 0x106
3c: 828a add.n a8, a2, a8
3e: ffff91 l32r a9, 3c <aot_func_internal#0+0x10> (ff91828a <aot_func_internal#0+0xff91825e>)
3e: R_XTENSA_SLOT0_OP .literal+0x8
41: 004892 s8i a9, a8, 0
44: 05a182 movi a8, 0x105
47: 828a add.n a8, a2, a8
49: ffff91 l32r a9, 48 <aot_func_internal#0+0x1c> (ffff9182 <aot_func_internal#0+0xffff9156>)
49: R_XTENSA_SLOT0_OP .literal+0xc
4c: 41a890 srli a10, a9, 8
4f: 0048a2 s8i a10, a8, 0
52: 04a182 movi a8, 0x104
55: 828a add.n a8, a2, a8
57: 004892 s8i a9, a8, 0
5a: f01d retw.n
```
Note that the each four bytes are stored separately using
one-byte-store instruction, s8i.
This commit tries to use larger alignments for load/store LLVM IR
instructions when possible. with this commit, the above example is
compiled into the following machine code, which seems more reasonable.
```
0000002c <aot_func_internal#0>:
2c: 004136 entry a1, 32
2f: ffff81 l32r a8, 2c <aot_func_internal#0> (81004136 <aot_func_internal#0+0x8100410a>)
2f: R_XTENSA_SLOT0_OP .literal+0x8
32: 416282 s32i a8, a2, 0x104
35: f01d retw.n
```
Note: this doesn't work well for --xip because aot_load_const_from_table()
hides the constness of the value. Maybe we need our own mechanism to
propagate the constness and the value.
Probably it's better to skip the optimized out parameters.
(that is, parameters w/o locations)
However, I'm not sure how/if it can be done with the lldb api.
For now, just disable parameter processing to avoid crashes.
While band-aid fixes like this is not plausible IMO,
some people prefer to have some debug info even if it's
partial/limited/broken. This commit partially (re)enables
C++ processing. On the other hand, do not bother to process
variables because it's known incompatible with C++.
Although I don't know what exactly the esp32s3 rom version of memset is,
it seems that the current code crashes only with a small "len". I guess
it changes the logic depending on the size.
Anyway, it's safer to use dbus here.
