Description
I recently encountered an unexpected behavior when compiling a WebAssembly module with wasm-ld using Position-Independent Executable (PIE) mode and Thread-Local Storage (TLS) (__thread variables). The generated code appears to access a string in .data at an incorrect offset, which causes misalignment.
I am unsure if this is an intended behavior or a potential bug in wasm-ld. I would appreciate any insights from the wasm-ld maintainers.
Reproduce the behavior
When compiling the following obj.c and lib.c files:
// obj.c
__thread int test = 10;
void obj_func() {
test = 1;
}
// lib.c
extern void obj_func();
int str_func(const char *s) {
return 42;
}
int lib_func() {
obj_func();
return str_func("this is a string");
}
using
clang --target=wasm32-unknown-wasi \
-nostdlib -fPIC -matomics -mbulk-memory -c obj.c -o obj.o
clang --target=wasm32-unknown-wasi \
-fPIC -nostdlib -Wl,-pie,--no-entry,--export=lib_func lib.c obj.o -o lib.wasm
I found that the generated WAT file places the string in .data with this definition:
(data $.data (global.get $__memory_base) "\0a\00\00\00this is a string\00")
which means the string starts at an offset of 4 relative to __memory_base. However, in the compiled code, the reference to the string is:
i32.const 8
local.set 0
global.get $__memory_base
local.set 1
local.get 1
local.get 0
i32.add
local.set 2
local.get 2
call $str_func
which uses an offset of 8 instead of 4.
And I also found out that If I remove __thread from obj.c, the issue disappears, and the correct offset is used.
I encountered this behavior when I was using clang version 18.1.8 with LLD version 18.1.8. I also tried to download the LLVM main branch and made a fresh build, and I am still encountering the same behavior.