* move inferred error sets into InternPool.
- they are now represented by pointing directly at the corresponding
function body value.
* inferred error set working memory is now in Sema and expires after
the Sema for the function corresponding to the inferred error set is
finished having its body analyzed.
* error sets use a InternPool.Index.Slice rather than an actual slice
to avoid lifetime issues.
Abridged summary:
* Move `Module.Fn` into `InternPool`.
* Delete a lot of confusing and problematic `Sema` logic related to
generic function calls.
This commit removes `Module.Fn` and replaces it with two new
`InternPool.Tag` values:
* `func_decl` - corresponding to a function declared in the source
code. This one contains line/column numbers, zir_body_inst, etc.
* `func_instance` - one for each monomorphization of a generic
function. Contains a reference to the `func_decl` from whence the
instantiation came, along with the `comptime` parameter values (or
types in the case of `anytype`)
Since `InternPool` provides deduplication on these values, these fields
are now deleted from `Module`:
* `monomorphed_func_keys`
* `monomorphed_funcs`
* `align_stack_fns`
Instead of these, Sema logic for generic function instantiation now
unconditionally evaluates the function prototype expression for every
generic callsite. This is technically required in order for type
coercions to work. The previous code had some dubious, probably wrong
hacks to make things work, such as `hashUncoerced`. I'm not 100% sure
how we were able to eliminate that function and still pass all the
behavior tests, but I'm pretty sure things were still broken without
doing type coercion for every generic function call argument.
After the function prototype is evaluated, it produces a deduplicated
`func_instance` `InternPool.Index` which can then be used for the
generic function call.
Some other nice things made by this simplification are the removal of
`comptime_args_fn_inst` and `preallocated_new_func` from `Sema`, and the
messy logic associated with them.
I have not yet been able to measure the perf of this against master
branch. On one hand, it reduces memory usage and pointer chasing of the
most heavily used `InternPool` Tag - function bodies - but on the other
hand, it does evaluate function prototype expressions more than before.
We will soon find out.
This adds a standalone test case to ensure the runtime does not trap
when performing a memory.copy or memory.fill instruction while the
destination or source address is out-of-bounds and the length is 0.
When lowering the `memset` instruction, LLVM will lower it to WebAssembly's
`memory.fill` instruction when the bulk-memory feature is enabled. This
instruction will trap when the destination address is out-of-bounds.
By Zig's semantics, it is valid to have an invalid pointer when the length is 0.
To prevent runtimes from trapping, we add a safety-check for slices to only
lower to a memset instruction when the length is larger than 0.
When lowering the `memcpy` instruction, LLVM will lower it to WebAssembly's
`memory.copy` instruction when the bulk-memory feature is enabled. This
instruction will trap when the destination or source pointer is out-of-bounds.
By Zig's semantics, it is valid to have an invalid pointer when the length is 0.
To prevent runtimes from trapping, we add a safety-check for slices to only
lower to a memcpy instruction when the length is larger than 0.
* `CMakeLists.txt`: support the weird `uname -m` output.
* `CMakeLists.txt`: detect and use the C compiler's default arm mode.
* cbe: support gcc with both `f128` and `u128` emulated.
* std.os.linux.thumb: fix incorrectly passed asm inputs.
The panic handler expects that this value is represented with the
generic address space, so cast the global to the generic address-
space before caching and returning the value.
Previously, interned values were represented as AIR instructions using
the `interned` tag. Now, the AIR ref directly encodes the InternPool
index. The encoding works as follows:
* If the ref matches one of the static values, it corresponds to the same InternPool index.
* Otherwise, if the MSB is 0, the ref corresponds to an InternPool index.
* Otherwise, if the MSB is 1, the ref corresponds to an AIR instruction index (after removing the MSB).
Note that since most static InternPool indices are low values (the
exceptions being `.none` and `.var_args_param_type`), the first rule is
almost a nop.
This actually used to be how it worked in stage1, and there was this
issue to change it: #2649
So this commit is a reversal to that idea. One motivation for that issue
was avoiding emitting the panic handler in compilations that do not have
any calls to panic. This commit only resolves the panic handler in the
event of a safety check function being emitted, so it does not have that
flaw.
The other reason given in that issue was for optimizations that elide
safety checks. It's yet to be determined whether that was a good idea or
not; this can get re-explored when we start adding optimization passes
to AIR.
This commit adds these AIR instructions, which are only emitted if
`backendSupportsFeature(.safety_checked_arithmetic)` is true:
* add_safe
* sub_safe
* mul_safe
It removes these nonsensical AIR instructions:
* addwrap_optimized
* subwrap_optimized
* mulwrap_optimized
The safety-checked arithmetic functions push the burden of invoking the
panic handler into the backend. This makes for a messier compiler
implementation, but it reduces the amount of AIR instructions emitted by
Sema, which reduces time spent in the secondary bottleneck of the
compiler. It also generates more compact LLVM IR, reducing time spent in
the primary bottleneck of the compiler.
Finally, it eliminates 1 stack allocation per safety-check which was
being used to store the resulting tuple. These allocations were going to
be annoying when combined with suspension points.
Most of this migration was performed automatically with `zig fmt`. There
were a few exceptions which I had to manually fix:
* `@alignCast` and `@addrSpaceCast` cannot be automatically rewritten
* `@truncate`'s fixup is incorrect for vectors
* Test cases are not formatted, and their error locations change
DeclGen/FuncGen methods are for things that pertain to a particular
declaration or function, while Object methods are for things that
pertain to the entire LLVM Module. Many methods were in the wrong
category, such as type and value lowering.
This is a prerequisite commit for a local branch I am working on, which
needs to be able to call lowerValue() without the context of any
particular function or declaration.
The signature is `getOrCreateSubrange(int64_t Lo, int64_t Count)`, so this updates the bindings to match.
This fixes a crash in `lowerDebugTypeImpl` when analyzing slices that have a length of 2^32 or
larger (up to `2^64 >> 3`, which still crashes, because above that the array size in bits overflows u64).
Anecdote 1: The generic version is way more popular than the non-generic
one in Zig codebase:
git grep -w alignForward | wc -l
56
git grep -w alignForwardGeneric | wc -l
149
git grep -w alignBackward | wc -l
6
git grep -w alignBackwardGeneric | wc -l
15
Anecdote 2: In my project (turbonss) that does much arithmetic and
alignment I exclusively use the Generic functions.
Anecdote 3: we used only the Generic versions in the Macho Man's linker
workshop.
These are frequently invalidated whenever a string is interned, so avoid
creating pointers to `string_bytes` wherever possible. This is an
attempt to fix random CI failures.
The main motivation for this commit is eliminating Decl.value_arena.
Everything else is dominoes.
Decl.name used to be stored in the GPA, now it is stored in InternPool.
It ended up being simpler to migrate other strings to be interned as
well, such as struct field names, union field names, and a few others.
This ended up requiring a big diff, sorry about that. But the changes
are pretty nice, we finally start to take advantage of InternPool's
existence.
global_error_set and error_name_list are simplified. Now it is a single
ArrayHashMap(NullTerminatedString, void) and the index is the error tag
value.
Module.tmp_hack_arena is re-introduced (it was removed in
eeff407941560ce8eb5b737b2436dfa93cfd3a0c) in order to deal with
comptime_args, optimized_order, and struct and union fields. After
structs and unions get moved into InternPool properly, tmp_hack_arena
can be deleted again.
This is neither a type nor a value. Simplifies `addStrLit` as well as
the many places that switch on `InternPool.Key`.
This is a partial revert of bec29b9e498e08202679aa29a45dab2a06a69a1e.
This is a bit odd, because this value doesn't actually exist:
see #15909. This gets all the empty enum/union behavior tests passing.
Also adds an assertion to `Sema.analyzeBodyInner` which would have
helped figure out the issue here much more quickly.
Key.PtrType is now an extern struct so that hashing it can be done by
reinterpreting bytes directly. It also uses the same representation for
type_pointer Tag encoding and the Key. Accessing pointer attributes now
requires packed struct access, however, many operations are now a copy
of a u32 rather than several independent fields.
This function moves the top two most used Key variants - pointer types
and pointer values - to use a single-shot hash function that branches
for small keys instead of calling memcpy.
As a result, perf against merge-base went from 1.17x ± 0.04 slower to
1.12x ± 0.04 slower. After the pointer value hashing was changed, total
CPU instructions spent in memcpy went from 4.40% to 4.08%, and after
additionally improving pointer type hashing, it further decreased to
3.72%.
The Zig language allows the compiler to make this optimization
automatically. We should definitely make the compiler do that, and
revert this commit. However, that will not happen in this branch, and I
want to continue to explore achieving performance parity with
merge-base. So, this commit changes all InternPool parameters to be
passed by const pointer rather than by value.
I measured a 1.03x ± 0.03 speedup vs the previous commit compiling the
(set of passing) behavior tests. Against merge-base, this commit is
1.17x ± 0.04 slower, which is an improvement from the previous
measurement of 1.22x ± 0.02.
Related issue: #13510
Related issue: #14129
Related issue: #15688
