Before this commit, GeneralPurposeAllocator could run into incredibly degraded performance in scenarios where the bucket count for a particular size class grew to be large. For example, if exactly `slot_count` allocations of a single size class were performed and then all of them were freed except one, then the bucket for those allocations would have to be kept around indefinitely. If that pattern of allocation were done over and over, then the bucket list for that size class could grow incredibly large.
This allocation pattern has been seen in the wild: https://github.com/Vexu/arocc/issues/508#issuecomment-1738275688
In that case, the length of the bucket list for the `128` size class would grow to tens of thousands of buckets and cause Debug runtime to balloon to ~8 minutes whereas with the c_allocator the Debug runtime would be ~3 seconds.
To address this, there are three different changes happening here:
1. std.Treap is used instead of a doubly linked list for the lists of buckets. This takes the time complexity of searchBucket [used in resize and free] from O(n) to O(log n), but increases the time complexity of insert from O(1) to O(log n) [before, all new buckets would get added to the head of the list]. Note: Any data structure with O(log n) or better search/insert/delete would also work for this use-case.
2. If the 'current' bucket for a size class is full, the list of buckets is never traversed and instead a new bucket is allocated. Previously, traversing the bucket list could only find a non-full bucket in specific circumstances, and only because of a separate optimization that is no longer needed (before, after any resize/free, the affected bucket would be moved to the head of the bucket list to allow searchBucket to perform better on average). Now, the current_bucket for each size class only changes when either (1) the current bucket is emptied/freed, or (2) a new bucket is allocated (due to the current bucket being full or null). Because each bucket's alloc_cursor only moves forward (i.e. slots within a bucket are never re-used), we can therefore always know that any bucket besides the current_bucket will be full, so traversing the list in the hopes of finding an existing non-full bucket is entirely pointless.
3. Size + alignment information for small allocations has been moved into the Bucket data instead of keeping it in a separate HashMap. This offers an improvement over the HashMap since whenever we need to get/modify the length/alignment of an allocation it's extremely likely we will already have calculated any bucket-related information necessary to get the data.
The first change is the most relevant and accounts for most of the benefit here. Also note that the overall functionality of GeneralPurposeAllocator is unchanged.
In the degraded `arocc` case, these changes bring Debug performance from ~8 minutes to ~20 seconds.
Benchmark 1: test-master.bat
Time (mean ± σ): 481.263 s ± 5.440 s [User: 479.159 s, System: 1.937 s]
Range (min … max): 477.416 s … 485.109 s 2 runs
Benchmark 2: test-optim-treap.bat
Time (mean ± σ): 19.639 s ± 0.037 s [User: 18.183 s, System: 1.452 s]
Range (min … max): 19.613 s … 19.665 s 2 runs
Summary
'test-optim-treap.bat' ran
24.51 ± 0.28 times faster than 'test-master.bat'
Note: Much of the time taken on Windows in this particular case is related to gathering stack traces. With `.stack_trace_frames = 0` the runtime goes down to 6.7 seconds, which is a little more than 2.5x slower compared to when the c_allocator is used.
These changes may or mat not introduce a slight performance regression in the average case:
Here's the standard library tests on Windows in Debug mode:
Benchmark 1 (10 runs): std-tests-master.exe
measurement mean ± σ min … max outliers delta
wall_time 16.0s ± 30.8ms 15.9s … 16.1s 1 (10%) 0%
peak_rss 42.8MB ± 8.24KB 42.8MB … 42.8MB 0 ( 0%) 0%
Benchmark 2 (10 runs): std-tests-optim-treap.exe
measurement mean ± σ min … max outliers delta
wall_time 16.2s ± 37.6ms 16.1s … 16.3s 0 ( 0%) 💩+ 1.3% ± 0.2%
peak_rss 42.8MB ± 5.18KB 42.8MB … 42.8MB 0 ( 0%) + 0.1% ± 0.0%
And on Linux:
Benchmark 1: ./test-master
Time (mean ± σ): 16.091 s ± 0.088 s [User: 15.856 s, System: 0.453 s]
Range (min … max): 15.870 s … 16.166 s 10 runs
Benchmark 2: ./test-optim-treap
Time (mean ± σ): 16.028 s ± 0.325 s [User: 15.755 s, System: 0.492 s]
Range (min … max): 15.735 s … 16.709 s 10 runs
Summary
'./test-optim-treap' ran
1.00 ± 0.02 times faster than './test-master'
This fixes a panic in `unionAbiSize` when a 0-length array of a union is used as a struct field.
Because `resolveTypeLayout` does not resolve the `elem_ty` if `arrayLenIncludingSentinel` returns
0 for the array, the child union type is not guaranteed to have a resolved layout at this point.
Fixed this case by just returning 0 here.
The SVG looks way better than the pixelated PNG and will adapt best to
whatever screen it is being displayed on. The PNG continues to be used
because Apple Safari does not support SVG favicons yet. All other major
browsers do. See https://caniuse.com/link-icon-svg.
This is a companion PR to ziglang/www.ziglang.org#310.
Notable changes in this update:
127198e58cb3dcf2d2287124cf15a23a7d3a9c02 fixes building zig2 artifact on
macOS Sonoma 14.0 (more specifically the SDK 14.0 linker).
a8d2ed806558cc1472f3a532169a4994abe17833 fixed some alignment edge
cases which is needed to do the store_hash=false change in the compiler
source code.
df5f0517b33b5f7bc2a508cf6a0ee62246f02d21 preserves result type
information through the address-of operator.
Instead of linear search every time a packed struct field's bit or byte
offset is wanted, they are computed once during resolution of the packed
struct's backing int type, and stored in InternPool for O(1) lookup.
Closes#17178
I have observed the standard library tests overflowing the default WASI
stack as of the previous commit. As best as I can tell, this isn't
directly our fault: LLVM is just emitting less efficient code in debug
builds with the new codegen patterns.
This commit introduces the new `ref_coerced_ty` result type into AstGen.
This represents a expression which we want to treat as an lvalue, and
the pointer will be coerced to a given type.
This change gives known result types to many expressions, in particular
struct and array initializations. This allows certain casts to work
which previously required explicitly specifying types via `@as`. It also
eliminates our dependence on anonymous struct types for expressions of
the form `&.{ ... }` - this paves the way for #16865, and also results
in less Sema magic happening for such initializations, also leading to
potentially better runtime code.
As part of these changes, this commit also implements #17194 by
disallowing RLS on explicitly-typed struct and array initializations.
Apologies for linking these changes - it seemed rather pointless to try
and separate them, since they both make big changes to struct and array
initializations in AstGen. The rationale for this change can be found in
the proposal - in essence, performing RLS whilst maintaining the
semantics of the intermediary type is a very difficult problem to solve.
This allowed the problematic `coerce_result_ptr` ZIR instruction to be
completely eliminated, which in turn also simplified the logic for
inferred allocations in Sema - thanks to this, we almost break even on
line count!
In doing this, the ZIR instructions surrounding these initializations
have been restructured - some have been added and removed, and others
renamed for clarity (and their semantics changed slightly). In order to
optimize ZIR tag count, the `struct_init_anon_ref` and
`array_init_anon_ref` instructions have been removed in favour of using
`ref` on a standard anonymous value initialization, since these
instructions are now virtually never used.
Lastly, it's worth noting that this commit introduces a slightly strange
source of generic poison types: in the expression `@as(*anyopaque, &x)`,
the sub-expression `x` has a generic poison result type, despite no
generic code being involved. This turns out to be a logical choice,
because we don't know the result type for `x`, and the generic poison
type represents precisely this case, providing the semantics we need.
Resolves: #16512Resolves: #17194
Source location resolution previously made ZIR printing incredibly slow,
since it was O(N^2). Since we usually resolve source locations
approximately in order, it is much more efficient to resolve them using
a "cursor" which navigates the file.
This takes the time for `zig ast-check -t Sema.zig` down from many
minutes (enough that I got bored and killed the process; well over 10)
to a few seconds.
SPIR-V doesn't support true element indexing, so we probably
need to switch over to isByRef like in llvm for this to work
properly. Currently a temporary is used, which at least
seems to work.
