A new `Legalize.Feature` tag is introduced for each float bit width
(16/32/64/80/128). When e.g. `soft_f16` is enabled, all arithmetic and
comparison operations on `f16` are converted to calls to the appropriate
compiler_rt function using the new AIR tag `.legalize_compiler_rt_call`.
This includes casts where the source *or* target type is `f16`, or
integer<=>float conversions to or from `f16`. Occasionally, operations
are legalized to blocks because there is extra code required; for
instance, legalizing `@floatFromInt` where the integer type is larger
than 64 bits requires calling an arbitrary-width integer conversion
function which accepts a pointer to the integer, so we need to use
`alloc` to create such a pointer, and store the integer there (after
possibly zero-extending or sign-extending it).
No backend currently uses these new legalizations (and as such, no
backend currently needs to implement `.legalize_compiler_rt_call`).
However, for testing purposes, I tried modifying the self-hosted x86_64
backend to enable all of the soft-float features (and implement the AIR
instruction). This modified backend was able to pass all of the behavior
tests (except for one `@mod` test where the LLVM backend has a bug
resulting in incorrect compiler-rt behavior!), including the tests
specific to the self-hosted x86_64 backend.
`f16` and `f80` legalizations are likely of particular interest to
backend developers, because most architectures do not have instructions
to operate on these types. However, enabling *all* of these legalization
passes can be useful when developing a new backend to hit the ground
running and pass a good amount of tests more easily.
I had tried unrolling the loops to avoid requiring the
`vector_store_elem` instruction, but it's arguably a problem to generate
O(N) code for an operation on `@Vector(N, T)`. In addition, that
lowering emitted a lot of `.aggregate_init` instructions, which is
itself a quite difficult operation to codegen.
This requires reintroducing runtime vector indexing internally. However,
I've put it in a couple of instructions which are intended only for use
by `Air.Legalize`, named `legalize_vec_elem_val` (like `array_elem_val`,
but for indexing a vector with a runtime-known index) and
`legalize_vec_store_elem` (like the old `vector_store_elem`
instruction). These are explicitly documented as *not* being emitted by
Sema, so need only be implemented by backends if they actually use an
`Air.Legalize.Feature` which emits them (otherwise they can be marked as
`unreachable`).
I started this diff trying to remove a little dead code from the C
backend, but ended up finding a bunch of dead code sprinkled all over
the place:
* `packed` handling in the C backend which was made dead by `Legalize`
* Representation of pointers to runtime-known vector indices
* Handling for the `vector_store_elem` AIR instruction (now removed)
* Old tuple handling from when they used the InternPool repr of structs
* Straightforward unused functions
* TODOs in the LLVM backend for features which Zig just does not support
This safety check was completely broken; it triggered unchecked illegal
behavior *in order to implement the safety check*. You definitely can't
do that! Instead, we must explicitly check the boundaries. This is a
tiny bit fiddly, because we need to make sure we do floating-point
rounding in the correct direction, and also handle the fact that the
operation truncates so the boundary works differently for min vs max.
Instead of implementing this safety check in Sema, there are now
dedicated AIR instructions for safety-checked intfromfloat (two
instructions; which one is used depends on the float mode). Currently,
no backend directly implements them; instead, a `Legalize.Feature` is
added which expands the safety check, and this feature is enabled for
all backends we currently test, including the LLVM backend.
The `u0` case is still handled in Sema, because Sema needs to check for
that anyway due to the comptime-known result. The old safety check here
was also completely broken and has therefore been rewritten. In that
case, we just check for 'abs(input) < 1.0'.
I've added a bunch of test coverage for the boundary cases of
`@intFromFloat`, both for successes (in `test/behavior/cast.zig`) and
failures (in `test/cases/safety/`).
Resolves: #24161
Runtime `@shuffle` has two cases which backends generally want to handle
differently for efficiency:
* One runtime vector operand; some result elements may be comptime-known
* Two runtime vector operands; some result elements may be undefined
The latter case happens if both vectors given to `@shuffle` are
runtime-known and they are both used (i.e. the mask refers to them).
Otherwise, if the result is not entirely comptime-known, we are in the
former case. `Sema` now diffentiates these two cases in the AIR so that
backends can easily handle them however they want to. Note that this
*doesn't* really involve Sema doing any more work than it would
otherwise need to, so there's not really a negative here!
Most existing backends have their lowerings for `@shuffle` migrated in
this commit. The LLVM backend uses new lowerings suggested by Jacob as
ones which it will handle effectively. The x86_64 backend has not yet
been migrated; for now there's a panic in there. Jacob will implement
that before this is merged anywhere.
This adds 4 `Legalize.Feature`s:
* `expand_intcast_safe`
* `expand_add_safe`
* `expand_sub_safe`
* `expand_mul_safe`
These do pretty much what they say on the tin. This logic was previously
in Sema, used when `Zcu.Feature.safety_checked_instructions` was not
supported by the backend. That `Zcu.Feature` has been removed in favour
of this legalization.
Each target can opt into different sets of legalize features.
By performing these transformations before liveness, instructions
that become unreferenced will have up-to-date liveness information.
Pointers to thread-local variables do not have their addresses known
until runtime, so it is nonsensical for them to be comptime-known. There
was logic in the compiler which was essentially attempting to treat them
as not being comptime-known despite the pointer being an interned value.
This was a bit of a mess, the check was frequent enough to actually show
up in compiler profiles, and it was very awkward for backends to deal
with, because they had to grapple with the fact that a "constant" they
were lowering might actually require runtime operations.
So, instead, do not consider these pointers to be comptime-known in
*any* way. Never intern such a pointer; instead, when the address of a
threadlocal is taken, emit an AIR instruction which computes the pointer
at runtime. This avoids lots of special handling for TLVs across
basically all codegen backends; of all somewhat-functional backends, the
only one which wasn't improved by this change was the LLVM backend,
because LLVM pretends this complexity around threadlocals doesn't exist.
This change simplifies Sema and codegen, avoids a potential source of
bugs, and potentially improves Sema performance very slightly by
avoiding a non-trivial check on a hot path.
Sema is arbitrarily scalarizing some operations, which means that when I
try to implement vectorized versions of those operations in a backend,
they are impossible to test due to Sema not producing them. Now, I can
implement them and then temporarily enable the new feature for that
backend in order to test them. Once the backend supports all of them,
the feature can be permanently enabled.
This also deletes the Air instructions `int_from_bool` and
`int_from_ptr`, which are just bitcasts with a fixed result type, since
changing `un_op` to `ty_op` takes up the same amount of memory.
This instruction is like `intcast`, but includes two safety checks:
* Checks that the int is in range of the destination type
* If the destination type is an exhaustive enum, checks that the int
is a named enum value
This instruction is locked behind the `safety_checked_instructions`
backend feature; if unsupported, Sema will emit a fallback, as with
other safety-checked instructions.
This instruction is used to add a missing safety check for `@enumFromInt`
truncating bits. This check also has a fallback for backends which do
not yet support `safety_checked_instructions`.
Resolves: #21946
The original motivation here was to fix regressions caused by #22414.
However, while working on this, I ended up discussing a language
simplification with Andrew, which changes things a little from how they
worked before #22414.
The main user-facing change here is that any reference to a prior
function parameter, even if potentially comptime-known at the usage
site or even not analyzed, now makes a function generic. This applies
even if the parameter being referenced is not a `comptime` parameter,
since it could still be populated when performing an inline call. This
is a breaking language change.
The detection of this is done in AstGen; when evaluating a parameter
type or return type, we track whether it referenced any prior parameter,
and if so, we mark this type as being "generic" in ZIR. This will cause
Sema to not evaluate it until the time of instantiation or inline call.
A lovely consequence of this from an implementation perspective is that
it eliminates the need for most of the "generic poison" system. In
particular, `error.GenericPoison` is now completely unnecessary, because
we identify generic expressions earlier in the pipeline; this simplifies
the compiler and avoids redundant work. This also entirely eliminates
the concept of the "generic poison value". The only remnant of this
system is the "generic poison type" (`Type.generic_poison` and
`InternPool.Index.generic_poison_type`). This type is used in two
places:
* During semantic analysis, to represent an unknown result type.
* When storing generic function types, to represent a generic parameter/return type.
It's possible that these use cases should instead use `.none`, but I
leave that investigation to a future adventurer.
One last thing. Prior to #22414, inline calls were a little inefficient,
because they re-evaluated even non-generic parameter types whenever they
were called. Changing this behavior is what ultimately led to #22538.
Well, because the new logic will mark a type expression as generic if
there is any change its resolved type could differ in an inline call,
this redundant work is unnecessary! So, this is another way in which the
new design reduces redundant work and complexity.
Resolves: #22494Resolves: #22532Resolves: #22538
Previously, stepping from the single statement within the loop would
always exit the loop because all of the code unrolled from the loop is
associated with the same line and treated by the debugger as one line.
This commit reworks how anonymous struct literals and tuples work.
Previously, an untyped anonymous struct literal
(e.g. `const x = .{ .a = 123 }`) was given an "anonymous struct type",
which is a special kind of struct which coerces using structural
equivalence. This mechanism was a holdover from before we used
RLS / result types as the primary mechanism of type inference. This
commit changes the language so that the type assigned here is a "normal"
struct type. It uses a form of equivalence based on the AST node and the
type's structure, much like a reified (`@Type`) type.
Additionally, tuples have been simplified. The distinction between
"simple" and "complex" tuple types is eliminated. All tuples, even those
explicitly declared using `struct { ... }` syntax, use structural
equivalence, and do not undergo staged type resolution. Tuples are very
restricted: they cannot have non-`auto` layouts, cannot have aligned
fields, and cannot have default values with the exception of `comptime`
fields. Tuples currently do not have optimized layout, but this can be
changed in the future.
This change simplifies the language, and fixes some problematic
coercions through pointers which led to unintuitive behavior.
Resolves: #16865
This commit introduces a new AIR instruction, `repeat`, which causes
control flow to move back to the start of a given AIR loop. `loop`
instructions will no longer automatically perform this operation after
control flow reaches the end of the body.
The motivation for making this change now was really just consistency
with the upcoming implementation of #8220: it wouldn't make sense to
have this feature work significantly differently. However, there were
already some TODOs kicking around which wanted this feature. It's useful
for two key reasons:
* It allows loops over AIR instruction bodies to loop precisely until
they reach a `noreturn` instruction. This allows for tail calling a
few things, and avoiding a range check on each iteration of a hot
path, plus gives a nice assertion that validates AIR structure a
little. This is a very minor benefit, which this commit does apply to
the LLVM and C backends.
* It should allow for more compact ZIR and AIR to be emitted by having
AstGen emit `repeat` instructions more often rather than having
`continue` statements `break` to a `block` which is *followed* by a
`repeat`. This is done in status quo because `repeat` instructions
only ever cause the direct parent block to repeat. Now that AIR is
more flexible, this flexibility can be pretty trivially extended to
ZIR, and we can then emit better ZIR. This commit does not implement
this.
Support for this feature is currently regressed on all self-hosted
native backends, including x86_64. This support will be added where
necessary before this branch is merged.
This commit modifies the representation of the AIR `switch_br`
instruction to represent ranges in cases. Previously, Sema emitted
different AIR in the case of a range, where the `else` branch of the
`switch_br` contained a simple `cond_br` for each such case which did a
simple range check (`x > a and x < b`). Not only does this add
complexity to Sema, which we would like to minimize, but it also gets in
the way of the implementation of #8220. That proposal turns certain
`switch` statements into a looping construct, and for optimization
purposes, we want to lower this to AIR fairly directly (i.e. without
involving a `loop` instruction). That means we would ideally like a
single instruction to represent the entire `switch` statement, so that
we can dispatch back to it with a different operand as in #8220. This is
not really possible to do correctly under the status quo system.
This commit implements lowering of this new `switch_br` usage in the
LLVM and C backends. The C backend just turns any case containing ranges
entirely into conditionals, as before. The LLVM backend is a little
smarter, and puts scalar items into the `switch` instruction, only using
conditionals for the range cases (which direct to the same bb). All
remaining self-hosted backends are temporarily regressed in the presence
of switch range cases. This functionality will be restored for at least
the x86_64 backend before merge.
