When converting values between two concrete types, handle the case that
the value is unrepresentable by the target type.
For integers, the converted value will be either the maximum or minimum
value for the integer type.
For floats, the converted value will be positive or negative infinity.
Bug: tint:1504
Change-Id: Ia56fb8170c0ea994632194f166062823d9507249
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91621
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
When abstract floats are supported, FloatLiteralExpression may hold values outside of the range of a f32. When we call FloatToString(), we may emit an 'inf', which is not a token for any backend.
Cast to f32 first, to ensure behavior remains consistent.
Note: Once the resolver materializes, and we constant fold, the backends shouldn't see any values outside of a f32.
Bug: tint:1504
Change-Id: I11942304a063d72302dad32e0d6d4e04aa39b5f8
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91425
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Instead of vector-of-variant.
This:
• Makes it impossible to produce a mix of scalar variant types, which
would make no sense.
• Reduces the size of a Constant, by removing the union-tag from each
element.
Also clean up terminology. Rename 'Constant::Scalar' to
'Constant::Element'. Scalars are well-defined in WGSL, and with the
introduction of abstract-numerics, this no longer makes sense.
Bug: tint:1504
Change-Id: I599aa97ad1ea798b7db8e512a5990ba75827faad
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91304
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Commit-Queue: Ben Clayton <bclayton@google.com>
• Add tint::CheckedConvert for converting between Number values and
checking that the value fits in the target type.
• Quantize the f16 values.
• Add tint::NumberUnwrapper<T> to obtain the underlying type of a
number.
• Add ostream '<<' operator.
• Add inequality operators.
Bug: tint:1504
Change-Id: I7afa64867a8df0e55ccee16de14ce6a93fbe1965
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91303
Reviewed-by: David Neto <dneto@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
These are the first intrinsics to support abstract-integers and
abstract-floats.
Nothing can use these yet, as the resolver hasn't had abstract numerics
enabled in production builds.
Bug: tint:1504
Change-Id: I2760f72f5d3c68c30c2e7de5ae5fb290dce64d3f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91027
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Attempting to emit shader code from an invalid program is unsafe.
The WGSL writer is used for dumping information which is valuable for debugging bad programs, so this has not been changed.
Bug: chromium:1327461
Change-Id: I4497fcb19d126ef6c872e2bcda8e9b79044aeb68
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91163
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
When enabled, unsuffixed literals will be treated as abstract numerics.
By having this disabled by default, we can build up the tests without
the risk of breaking production code.
This flag is enabled for resolver unit tests. This can safely be done
with no change in tested behavior, as all tests use the '_i' or '_u'
literal suffixes, so currently there are no tests that exercise abstract
numerics.
Bug: tint:1504
Change-Id: I39523ff6e235a12533b1dd98587b580bed98300f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91025
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
• Add sem::SwitchStatement::Cases()
• Add sem::CaseStatement::Selectors()
• Add ast::SwitchStatement -> sem::SwitchStatement mapping
Removes a bunch of hopping between the AST and SEM to get at this data.
Change-Id: If48d78e7a386aa0b34c6d00ad9af1d53cb236f12
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91024
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
• Add Name() method so we can print the type name without having to
create a ProgramBuilder.
• Add a DataType for f16.
• Add AST() methods on DataType<AInt> and DataType<AFloat> which return
nullptr.
Change-Id: I130648655e114c658a0c2985822630a71a683dd6
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91023
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
These two functions in HLSL only accept and return uint. Thus, if the
result of these calls is passed to a function that wants int, it will
fail, or call the uint overload if one exists. Fixed by casting the
result to int if the arg is int.
Bug: tint:1550
Change-Id: Id4c0970a29ac4c83ee5b78be8d2762e05e4a3f03
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91001
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Reviewed-by: Ryan Harrison <rharrison@chromium.org>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
• Add sem::Type::Common() which returns the 'common' type for the list of
types.
• Migrate intrisnic table to use this.
• Add a whole-lotta-tests.
• Deduplicate and improve the EXPECT_TEST() macro. Move it to a common home.
Bug: tint:1504
Change-Id: I1564f67ecf87fc594f3f54274da906ff0d822795
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/91020
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: David Neto <dneto@google.com>
With abstract materialization, any ast::Expression may map to the new
sem::Materialize node. Because of this, we can't assume that an
ast::CallExpression maps to a sem::Call, as it might be wrapped by a
sem::Materialize.
Remove the mapping, and fix up all the code that was relying on this.
Fixes are done by either:
• Calling `UnwrapMaterialize()->As<sem::Call>()` on the semantic
expression. This is done when the logic may assume it's possible for
the expression to be a Materialize node.
• Using the explicit sem::Info::Get<sem::Call>() template argument to
cast the semantic type to sem::Call. This is done when the logic
either knows it is impossible for the expression to be a Materialize.
The backends have been stubbed, as we'll want to emit the constant value
for these nodes. It's likely that we'll just use the FoldConstants
transform to strip all Materialize nodes from the tree. For now, be
defensive.
Bug: tint:1504
Change-Id: If9231b300fc30c7fe886c17a804ead8ee2988285
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90533
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
• Split type_ctors_ into array_ctors_ and struct_ctors_. These are the
only two type constructors that we need to cache in the Resovler. All
other types are cached in the IntrinsicTable. By making these two
separate fields, there's no way we can accidentally put other types in
there. This is slightly more optimized too, as each map is smaller,
and the keys hold less data.
• Drop the 'const' on vectors that are std::move()'d. A const object
cannot be moved, and this results in a silent copy. Fix the logic of
BuiltinCall() and FunctionCall(), as these were happily using the
vectors that had been moved.
• Extract the messy Texture-Sampling collection logic out into two
functions.
The rest of the changes are tweaks required to handle abstract numerics
and materialization, which will be put up for review shortly. Landing
the no-op refactoring now reduces review noise later.
Bug: tint:1504
Change-Id: Iffc8039360d6138c3ac9b456be6ca7b8451ede9f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90532
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
A new semantic expression node that wraps another semantic node. Used to
indicate the point at which compile-time, abstract numeric typed
expressions are implicitly converted to a concrete type.
Bug: tint:1504
Change-Id: I52e256bbbdeaa9d9eff4cb93b6f937dd00bdc5cb
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90531
Reviewed-by: David Neto <dneto@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
TypeMatcher::String() was not respecting the [[display]] decoration of
the matcher's sub-types. By calling TypeMatcher::String() on the sub-types,
we can display the custom type names in diagnostics.
Bug: tint:1504
Change-Id: I0856fee31231f9c048d2e3028d25c4d261fbb008
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90529
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
* Extract ast::Enable::ExtensionKind to ast::Extension.
* Move the parsing out of ast::Enable and next to ast/extension.h
* Change the ast::Enable constructor to take the Extension, instead of
a std::string. It's the WGSL parser's responsibility to parse, not the
AST nodes.
* Add ProgramBuilder::Enable() helper.
* Keep ast::Module simple - keep track of the declared AST Enable nodes,
don't do any deduplicating of the enabled extensions.
* Add the de-duplicated ast::Extensions to the sem::Module.
* Remove the kInternalExtensionForTesting enum value - we have kF16
now, which can be used instead for testing.
* Rename kNoExtension to kNone.
Bug: tint:1472
Change-Id: I9af635e95d36991ea468e6e0bf6798bb50937edc
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90523
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Support overload resolution of abstract-numeric argument types,
allowing them to implicitly convert down to concrete parameter
types (and in the near future, abstract parameter types).
Major kudos to cwallez for the suggested algorithm which is a
minor adaption of what we had already.
Bug: tint:1504
Change-Id: I85fa8e70ab0b6aa643caec4c51433f15784af55f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90522
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Add support for a [[precedence(N)]] decoration on intrinsic table type
declarations. This will be used to ensure the type with the lowest
conversion rank is matched when a matcher could match multiple types
for a given abstract numeric argument type.
Bug: tint:1504
Change-Id: I96475b000c0917bbfa4e2873b1731ce048b96a7d
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90664
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
The concept of 'closing' an open type or number made sense when these
were immutable once the first type/number had been matched during
overload resolution.
In order to support abstract numerics, these template parameters need to
be constrained as the arguments are evaluated, so there's no longer a
binary open / closed state.
Give up on this concept, and rename everything to 'template type' and
'template number'. This is likely easier for people to understand
anyway.
Also fix a small typo in the ICE message printed when there's an
ambiguous overload resolution (should never happen with the current
entries in the table).
Bug: tint:1504
Change-Id: I2bf043c71e5afa757259968eae4af830c50f38e0
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90662
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: David Neto <dneto@google.com>
Migrate from a hand-rolled tagged-union of [i32, u32, f32, f16, bool]
types. Instead use a std::variant of [AInt, AFloat, bool]. The Constant
holds the actual type, so no information is lost with the reduced types.
Note: Currently integer constants are still limited to 32-bits in size.
This is enforced by the frontend.
Bug: tint:1504
Change-Id: I316957787649c454fffb532334159d726cd1fb2d
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90643
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Remove the ProgramBuilder from ClosedState and use a pointer for the
'overload' field instead of a reference. Let's the Candidate be
copy-assignable, which in turn, allows the Candidates vector to be
sorted directly, instead of jumping through hoops to use moves.
Replace random mix of 'int', 'uint8_t' with 'size_t' (externally to the
constant table data). Reduces fragile weak binding between distant code.
Swap the overload scoring order (high-best -> low-best). Remove the
'matched' field - we can now just check whether the 'score' is 0.
Further simplifies sorting.
Change-Id: I4a4b7934be337306202647d096c546eab5c8498f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90641
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
If the CallSiteRequiredToBeUniform tag is present, add the edge from
RequiredToBeUniform to a new diagnostic node for the function call,
instead of to the control flow coming out of the function call. Doing
the latter causes a false positive when a function both requires
uniform control flow and causes non-uniform control flow.
Bug: tint:880
Change-Id: Icade8f76302e8c21529502f5f945f1981acfc45a
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/90582
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Auto-Submit: James Price <jrprice@google.com>
