Revision 0d54ad3086b7fc61afa28b512b27668e1ddef2f5 authored by Keno Fischer on 17 April 2024, 23:01:19 UTC, committed by Keno Fischer on 17 April 2024, 23:40:48 UTC
The strategy here is to look at (data, padding) pairs and RLE them into loops, so that repeated adjacent patterns use a loop rather than getting unrolled. On the test case from #54109, this makes compilation essentially instant, while also being faster at runtime (turns out LLVM spends a massive amount of time AND the answer is bad). There's some obvious further enhancements possible here: 1. The `memcmp` constant is small. LLVM has a pass to inline these with better code. However, we don't have it turned on. We should consider vendoring it, though we may want to add some shorcutting to it to avoid having it iterate through each function. 2. This only does one level of sequence matching. It could be recursed to turn things into nested loops. However, this solves the immediate issue, so hopefully it's a useful start. Fixes #54109.
1 parent 7ba1b33
NEWS.md
Julia v1.12 Release Notes
========================
New language features
---------------------
Language changes
----------------
- When methods are replaced with exactly equivalent ones, the old method is no
longer deleted implicitly simultaneously, although the new method does take
priority and become more specific than the old method. Thus if the new
method is deleted later, the old method will resume operating. This can be
useful to mocking frameworks (such as in SparseArrays, Pluto, and Mocking,
among others), as they do not need to explicitly restore the old method.
While inference and compilation still must be repeated with this, it also
may pave the way for inference to be able to intelligently re-use the old
results, once the new method is deleted. ([#53415])
- Macro expansion will no longer eargerly recurse into into `Expr(:toplevel)`
expressions returned from macros. Instead, macro expansion of `:toplevel`
expressions will be delayed until evaluation time. This allows a later
expression within a given `:toplevel` expression to make use of macros
defined earlier in the same `:toplevel` expression. ([#53515])
Compiler/Runtime improvements
-----------------------------
- Generated LLVM IR now uses actual pointer types instead of passing pointers as integers.
This affects `llvmcall`: Inline LLVM IR should be updated to use `i8*` or `ptr` instead of
`i32` or `i64`, and remove unneeded `ptrtoint`/`inttoptr` conversions. For compatibility,
IR with integer pointers is still supported, but generates a deprecation warning. ([#53687])
Command-line option changes
---------------------------
* The `-m/--module` flag can be passed to run the `main` function inside a package with a set of arguments.
This `main` function should be declared using `@main` to indicate that it is an entry point.
Multi-threading changes
-----------------------
Build system changes
--------------------
New library functions
---------------------
* `logrange(start, stop; length)` makes a range of constant ratio, instead of constant step ([#39071])
* The new `isfull(c::Channel)` function can be used to check if `put!(c, some_value)` will block. ([#53159])
* `waitany(tasks; throw=false)` and `waitall(tasks; failfast=false, throw=false)` which wait multiple tasks at once ([#53341]).
New library features
--------------------
* `invmod(n, T)` where `T` is a native integer type now computes the modular inverse of `n` in the modular integer ring that `T` defines ([#52180]).
* `invmod(n)` is an abbreviation for `invmod(n, typeof(n))` for native integer types ([#52180]).
* `replace(string, pattern...)` now supports an optional `IO` argument to
write the output to a stream rather than returning a string ([#48625]).
* `sizehint!(s, n)` now supports an optional `shrink` argument to disable shrinking ([#51929]).
* New function `Docs.hasdoc(module, symbol)` tells whether a name has a docstring ([#52139]).
* New function `Docs.undocumented_names(module)` returns a module's undocumented public names ([#52413]).
* Passing an `IOBuffer` as a stdout argument for `Process` spawn now works as
expected, synchronized with `wait` or `success`, so a `Base.BufferStream` is
no longer required there for correctness to avoid data races ([#52461]).
* After a process exits, `closewrite` will no longer be automatically called on
the stream passed to it. Call `wait` on the process instead to ensure the
content is fully written, then call `closewrite` manually to avoid
data-races. Or use the callback form of `open` to have all that handled
automatically.
* `@timed` now additionally returns the elapsed compilation and recompilation time ([#52889])
* `filter` can now act on a `NamedTuple` ([#50795]).
* `tempname` can now take a suffix string to allow the file name to include a suffix and include that suffix in
the uniquing checking ([#53474])
* `RegexMatch` objects can now be used to construct `NamedTuple`s and `Dict`s ([#50988])
Standard library changes
------------------------
#### StyledStrings
#### JuliaSyntaxHighlighting
#### Package Manager
#### LinearAlgebra
#### Logging
#### Printf
#### Profile
#### Random
#### REPL
#### SuiteSparse
#### SparseArrays
#### Test
#### Dates
#### Statistics
#### Distributed
#### Unicode
#### DelimitedFiles
#### InteractiveUtils
Deprecated or removed
---------------------
External dependencies
---------------------
Tooling Improvements
--------------------
<!--- generated by NEWS-update.jl: -->
Computing file changes ...