While malloc_stats and friends are disabled and unused for wasi-libc,
it's neater to be consistent.
Background: My colleagues for some reasons enabled malloc_stats
and asked me why it reports negative values.
Note: Depending __heap_base, init_top() adjusts the address for
alignment. I think the amount of this adjustment is reported as
"used" by malloc_stats. I don't bother to "fix" it.
In a multi-threaded execution we need to make sure that only exactly one
thread initializes malloc. The function try_init_allocator() can't
easily be made thread-safe, so just move the call to
try_init_allocator() inside the block that holds the lock.
This commit effectively drops the support of older wasm-ld. (LLVM <15.0.7).
We have two relevant use cases:
* `memory.grow` use outside of malloc
(eg. used by polyfill preview1 binaries)
* `--init-memory` to somehow preallocate heap
(eg. avoid dynamic allocations, especially on small environments)
While https://github.com/WebAssembly/wasi-libc/pull/377
fixed the former, it broke the latter if you are using
an older LLVM, which doesn't provide the `__heap_end` symbol,
to link your module.
As we couldn't come up with a solution which satisfies all parties,
this commit simply makes it require new enough LLVM which provides
`__heap_end`. After all, a link-time failure is more friendly to users
than failing later in a subtle way.
* Don't use sbrk(0) to determine the initial heap size
This commit changes the `try_init_allocator` function as part of
dlmalloc to not use `sbrk(0)` to determine the initial heap size. The
purpose of this function is to use the extra memory at the end of linear
memory for the initial allocation heap before `memory.grow` is used to
allocate more memory. To learn the extent of this region the code
previously would use `sbrk(0)` to find the current size of linear
memory. This does not work, however, when other systems have called
`memory.grow` before this function is called. For example if another
allocator is used or if another component of a wasm binary grows memory
for its own purposes then that memory will be incorrectly claimed to be
owned by dlmalloc.
Instead this commit rounds up the `__heap_base` address to the nearest
page size, since that must be allocatable. Otherwise anything above this
rounded address is assumed to be used by something else, even if it's
addressable.
* Use `__heap_end` if defined
* Move mstate initialization earlier
* Remove -fno-builtin.
-fno-builtin suppresses optimizations such as turning calls to `sqrt`
or `fabs` into `f64.sqrt` or `f64.abs` instructions inline. Libc code
itself benefits from these optimizations.
I originally added this flag because historically it was needed when
building libc to avoid the compiler pattern-matching the body of memcpy
into a memcpy call, however clang no longer requires this.
* Expand the comment about why we use USE_DL_PREFIX in dlmalloc.
This isn't strictly required, as wasm SIMD loads and stores work on
unaligned memory. However, it may provide better performance. That said,
this isn't currently studied by any benchmarking.