| Age | Commit message (Collapse) | Author |
|---|
|
* Prefixing source files in source/slang with slang-
* Prefix source in source/slang with slang- prefix.
* Rename core source files with slang- prefix.
* Update project files.
* Fix problems from automatic merge.
|
|
Fixes #858
The `precise` keyword exists in both HLSL and GLSL and when applied to a variable declaration is supposed to indicate that all computations that contribute to the value of that variable should not be altered based on "fast-math" optimizations. The main examples are that separate multiply and add operations should not be turned into fused multiply-add (fma) operations, and that operations cannot ignore the possibility of infinity or not-a-number values (e.g., by assuming that `x * 0.0f` is always `0.0f`).
(Aside: it is possible that my understanding of what the semantics of `precise` are in HLSL and GLSL is imperfect so that either the GLSL variant isn't sufficient to provide the semantics of the HLSL keyword, or that the definition of "all computations that contribute" to a value isn't actually correct. We may need to revise this implementation based on subsequent learnings.)
The basic idea here is to turn the AST `precise` keyword into a `[precise]` decoration in the IR and then emit that as a `precise` keyword again in the output.
The main catch is that whereas most of our existing IR decorations apply to things like global shader parameters or `struct` members that usually stick around for the duration of compilation, `[precise]` will get slapped on local variables that will often get optimized away by our SSA pass. There are two ways a variable can get eliminated/replaced during the SSA pass:
1. A use of the variable can be replaced with an ordinary instruction that computes its value.
2. A use of the variable can be replaced with a reference to a "phi node" that will take on the appropriate value based on control flow.
These two cases already had logic to copy a "name hint" decoration from the variable over to an instruction that will replace it, and I simply extended them to also propagate over a `[precise]` decoration.
The test case added with this change intentionally constructs a case where `[precise]` needs to be propagated over to an SSA "phi node" in order to generate correct output code.
The other gotcha is that we can emit variable declarations in various places in `emit.cpp`, and all of these needed to handle `[precise]`. Not only do we have actually local variables (`IRVar`), but we also have SSA phi nodes (`IRParam`), and then there are cases where an intermediate computation (an ordinary instruction) should be `[precise]` and thus we need to emit it as a temporary (not folding it into its use sites) and make sure that the temporary itself gets the `precise` keyword.
I have manually confirmed that in the output SPIR-V, this change results in the `NoContraction` SPIR-V decoration being added to the relevant operations, and the output DXBC contains a multiply and an add in place of a multiply-add. The output DXIL does not show any obvious changes due to `precise`, although the exact order and operands of the math instructions emitted does differ when `precise` is added/removed. In all cases the output is equivalent to hand-written HLSL/GLSL with a `precise`-qualified local variable.
|
|
* Add better control over image formats for GLSL/SPIR-V targets
Currently Slang emits GLSL code assuming all R/W images need to have explicit formats, and thus we try to infer a format from the element type of the image.
E.g., given a `RWTexture2D<half4>` we might infer that a qualifier of `layout(rgba16f)` should be used.
This strategy has two notable shortcomings:
* Sometimes the user will want a format that doesn't match an existing HLSL type. E.g., if they want the equivalent of `layout(r11f_g11f_b10f)`, then what should they put in their `RWTexture2D<...>` to make the inference do what they need?
* Sometimes the user knows that they don't need to specify a format *at all*, because using the `GL_EXT_shader_image_load_formatted` extension, they can still perform non-atomic load/store on images with no format specified in the SPIR-V.
This change adds two features directed at these challenges.
First, we add an explicit `[format(...)]` attribute that can be used to specify an explicit image format, including ones that don't match any HLSL type.
An example of using this new attribute is:
```hlsl
[format("r11f_g11f_b10f")]
RWTexture2D<float3> myImage;
```
For simplicity in initial bring-up, the new formats all use the same naming as formats in GLSL (this should make it easy for a programmer who knows what they expect to get in the GLSL output). We can change the naming convention for formats at a later time, so long as we keep these existing names in as a compatibility feature.
Note that this is *not* given a `vk::` prefix since the attribute should signal the programmer's intent to provide an image with that format on *all* targets (although only some targets might act on that information).
Also note that the attribute takes a string (`[format("rgba8")`) instead of a bare identifier (`[format(rgba8)]`) because this is consistent with the existing convention for attributes in HLSL.
When `[format(...)]` is left off, the default compiler behavior will still be to infer a format, but this behavior can be overidden for a single image using an explicit format of `"unknown"`:
```hlsl
[format("unknown")]
RWTexture2D<float4> mysteryMachine;
```
The second new feature is that if a user knows they are coding for a GPU that supports the `"unknown"` format in all non-atomic cases, then they can opt into making that the default for images without an explicit `[format(...)]`, using the new `-default-image-format-unknown` command-line option for `slangc`.
The new test case included with this change confirms that we correctly see the explicit formats in the output GLSL and *no* formats for images without explicit `[format(...)]` when using the new command-line option. The test stresses images declared at global scope, in parameter blocks, and in entry-point parameter lists, to try and make sure that all the relevant IR passes in the compiler preserve the format information.
* fixup: missing file
|
|
* Improve support for interfaces as shader parameters
This change adds two main things over the existing support:
1. It is now possible to plug in concrete types that actually contain (uniform/ordinary) fields for the existential type parameters introduced by interface-type shader parameters. The `interface-shader-param2.slang` test shows that this works.
2. There is a limited amount of support for doing correct layout computation and generating output code that matches that layout, so that interface and ordinary-type fields can be interleaved to a limited extent. The `interface-shader-param3.slang` test confirms this behavior.
There are several moving pieces in the change.
* When it comes to terminology, we try to draw a more clear distinction between existial type parameters/arguments and existential/object value parametes/arguments. A simple way to look at it is that an `IFoo[3]` shader parameter introduces a single existential type parameter (so that a concrete type argument like `SomeThing` can be plugged in for the `IFoo`) but introduces three existential object/value parameters (to represent the concrete values for the array elements).
* At the IR level, we support a few new operations. A `BindExistentialsType` can take a type that is not itself an interface/existential type but which depends on interfaces/existentials (e.g., `ConstantBuffer<IFoo>`) and plug in the concrete types to be used for its existential type slots.
* Then a `wrapExistentials` instruction can take a type with all the existentials plugged in (possibly by `BindExistentialsType`) and wrap it into a value of the existential-using type (e.g., turn `ConstantBuffer<SomeThing>` into a `ConstantBuffer<IFoo>`).
* The IR passes for doing generic/existential specialization have been updated to be able to desugar uses of these new operations just enough so that a `ConstantBuffer<IFoo>` can be used.
* When we specialize an IR parameter of an interface type like `IFoo` based on a concrete type `SomeThing`, we turn the parameter into an `ExistentialBox<SomeThing>` to reflect the fact that we are conceptually referring to `SomeThing` indirectly (it shouldn't be factored into the layout of its surrounding type).
* Parameter binding was updated so that it passes along the bound existential type arguments in a `Program` or `EntryPoint` to type layout, so that we can take them into account. The type layout code needs to do a little work to pass the appropriate range of arguments along to sub-fields when computing layout for aggregate types.
* Type layout was updated to have a notion of "pending" items, which represent the concrete types of data that are logically being referenced by existential value slots. The basic idea is that these values aren't included in the layout of a type by default, but then they get "flushed" to come after all the non-existential-related data in a constant buffer, parameter block, etc.
* The logic for computing a parameter group (`ConstantBuffer` or `ParameterBlock`) layout was updated to always "flush" the pending items on the element type of the group, so that the resource usage of specialized existential slots would be taken into account.
* The type legalization pass has been adapted so that we can derive two different passes from it. One does resource-type legalization (which is all that the original pass did). The new pass uses the same basic machinery to legalize `ExistentialBox<T>` types by moving them out of their containing type(s), and then turning them into ordinary variables/parameters of type `T`.
Big things missing from this change include:
- Nothing is making sure that "pending" items at the global or entry-point level will get proper registers/bindings allocated to them. For the uniform case, all that matters in the current compiler is that we declare them in the right order in the output HLSL/GLSL, but for resources to be supported we will need to compute this layout information and start associating it with the existential/interface-type fields.
- Nothing is being done to support `BindExistentials<S, ...>` where `S` is a `struct` type that might have existential-type fields (or nested fields...). Eventually we need to desugar a type like this into a fresh `struct` type that has the same field keys as `S`, but with fields replaced by suitable `BindExistentials` as needed. (The hard part of this would seem to be computing which slots go to which fields). As a practial matter, this missing feature means that interface-type members of `cbuffer` declarations won't work.
The current tests carefully avoid both of these problems. They don't declare any buffer/texture fields in the concrete types, and they don't make use of `cbuffer` declarations or `ConstantBuffer`s over structure types with interface-type fields.
* fixup: add override to methods
* fixup: typos
|
|
* First steps toward supporting interface-type parameters on shaders
What's New
----------
From the perspective of a user, the main thing this change adds is the ability to declare top-level shader parameters (either at global scope, or in an entry-point parameter list) with interface types. For example, the following becomes possible:
```hlsl
// Define an interface to modify values
interface IModifier { float4 modify(float4 val); }
// Define some concrete implementations
struct Doubler : IModifier
{
float4 modify(float4 val) { return val + val; }
}
struct Squarer : IModifier { ... }
// Define a global shader parameter of interface type
IModifier gGlobalModifier;
// Define an entry point with an interface-type `uniform` parameter
void myShader(
unifrom IModifier entryPointModifier,
float4 inColor : COLOR,
out float4 outColor : SV_Target)
{
// Use the interface-type parameters to compute things
float4 color = inColor;
color = gGlobalModifier.modify(color);
color = entryPointModifier.modify(color);
outColor = color;
}
```
The user can specialize that shader by specifying the concrete types to use for global and entry-point parameters of interface types (e.g., plugging in `Doubler` for `gGlobalModifier` and `Squarer` for `entryPointModifier`).
The "plugging in" process is done in terms of a concept of both global and local "existential slots" which are a new `LayoutResourceKind` that represents the holes where concrete types need to be plugged in for existential/interface types.
In simple cases like the above, each interface-type parameter will yield a single existential slot in either the global or entry-point parameter layout. Users can query the start slot and number of slots for each shader parameter, just like they would for any other resource that a parameter can consume. Before generating specialized code, the user plugs in the name of the concrete type they would like to use for each slot using `spSetTypeNameForGlobalExistentialSlot` and/or `spSetTypeNameForEntryPointExistentialSlot`.
There are some major limitations to the implementation in this first change:
* Parameters must be of interface type (e.g., `IFoo`) and not an array (`IFoo[3]`), or buffer (`ConstantBuffer<IFoo>`) over an interface type. Similarly, `struct` types with interface-type fields still don't work.
* The work on interface-type function parameters still doesn't include support for `out` or `inout` parameters, nor for functions that return interface types (that isn't technically related to this change, but affects its usefullness).
* No work is being done to correctly lay out shader parameters once the concrete types for existential slots are known, so that this change really only works when the concrete type that gets plugged in is empty.
These limitations are severe enough that this feature isn't really usable as implemented in this change, and this merely represents a stepping stone toward a more complete implementation.
Implementation
--------------
The API side of thing largely mirrors what was already done to support passing strings for the type names to use for global/entry-point generic arguments, so there should be no major surprises there.
The logic in `check.cpp` computes the list of existential slots when creating unspecialized `Program`s and `EntryPoint`s (this is logically the "front end" of the compiler), and then checks the supplied argument types against what is expected in each slot when creating specialized `Program`s and `EntryPoint`s. This again mirrors how generic arguments are handled.
Type layout was extended to compute the number of existential slots that a type consumes, and will thus automatically assign ranges of slots to top-level and entry-point shader parameters in the same way it already allocates `register`s and `binding`s. The big missing feature is the ability to specialize a layout to account for the concrete types plugged into the existential-type slots.
IR generation for specialized programs and entry points was slightly extended so that it attaches information about the concrete types plugged into the existential slots, and the witness tables that show how they conform to the interface for that slot. The linking step needed some small tweaks to make sure that information gets copied over to the target-specific program when we start code generation.
The meat of the IR-level work is in `ir-bind-existentials.cpp`, which takes the information that was placed in the IR module by the generation/linking steps and uses it to rewrite shader parameters. For example, if there is a shader parameter `p` of type `IModifier`, and the corresponding existential slot has the type `Doubler` in it, we will rewrite the parameter to have type `Doubler`, and rewrite any uses of `p` to instead use `makeExistential(p, /*witness that Doubler conforms to IModifier*/)`.
Once the replacement is done on the parameters, the existing work for specializing existential-based code when the input type(s) are known kicks in and does the rest.
Testing
-------
A single compute test is added to validate that this feature works. It is narrowly tailored to not require any of the features not supported by the initial implementation (e.g., all of the concrete types used have no members).
The test case *does* include use of an associated type through one of these existential-type parameters, which has exposed a subtle bug in how "opening" of existential values is implemented in the front-end. Rather than fix the underlying problem, I cleaned up the code in the front-end to special case when the existential value being opened is a variable bound with `let`, to directly use a reference to that variable rather than introduce a temporary. Similarly, in the IR generation step, I added an optimization to make variables declared with `let` skip introducing an IR-level variable and just use the SSA value of their initializer directly instead.
* fixup: missing files
* fixup: incorrect type for unreachable return
* fixup: actually comment ir-bind-existentials.cpp
|
|
* Split front- and back-ends
This change is a major refactor of several of the types that provide the behind-the-scenes implementation of the public C API.
The goal of this refactor is primarily to allow for future API services that let the user operate both the front- and back-ends of the compiler in a more complex fashion.
For example, as user should be able to compile a bunch of source code into modules, look up types, functions, etc. in those modules, specialize generic types/functions to the types they've looked up, and then finally request target code to be gernerated for specialized entry points.
The back-end code generation they trigger should re-use the front-end compilation work (parsing, semantic checking, IR generation) that was already performed.
The most visible change is that `CompileRequest` has been split up into several smaller types that take responsibility for parts of what it did:
* The `Linkage` type owns the storage for `import`ed modules, and well as the `TargetRequest`s that represent code-generation targets. The intention is that an application could use a single `Linkage` for the duration of its runtime (so long as it was okay with the memory usage), so that each `import`ed module only gets loaded once. For now, this type needs to manage the search paths, file system, and source manager, because of its responsibility for loading files.
* A `FrontEndCompileRequest` owns the stuff related to parsing, semantic checking, and initial IR generation. This most notably includes the `TranslationUnitRequest`s and the `FrontEndEntryPointRequest`s (which used to be just `EntryPointRequest`s). It's main job is to produce AST and IR modules for each translation unit, and to find and validate the entry points. The front-end request does *not* interact with generic arguments for global or entry-point generic parameters.
* The main output of both `import` operations and front-end translation units is the `Module` type, which is just a simple container for both the AST module (to service the reflection/layout APIs, and also for semantic checking of code that `import`s the module) and the IR module (for linking and code generation). This type captures the commonalities between the old `LoadedModule` (which is now just an alias for `Module`) and `TranslationUnitRequest` (which now owns a `Module`).
* The secondary output of front-end compilation is a `Program`, which comprises a list of referenced `Module`s and validated `EntryPoint`s that will be used together. Layout and code generation both need a `Program` to tell them what modules and entry points will be used together (we don't want to just code-gen everythin that has ever been loaded into the linakge). The `Program`s created by the front-end do not include generic arguments, so they may provide incomplete layout information and/or be unsuitable for code generation.
* A `BackEndCompileRequest` owns stuff related to turning a `Program` into output kernels for the targets of a `Linkage`. Most of the data it owns beyond the `Program` to be compiled is minor, so this is a good candidate for demotion from a heap-allocated object to just a `struct` of options that gets passed around.
* The `CompileRequestBase` type is an attempt to wrap up the common functionality of both front-end and back-end compile requests. Most of it is just exposing the availability of a linkage and `DiagnosticSink`, so this type is a good candidate for subsequent removal. The main interesting thing it has is the flags related to dumping and validation of IR, so there is probably a good refactoring still to be made around deciding how options should be handled going forward.
* Behind the scenes, the `Program` type is set up to handle some level of on-line compilation and layout work. The `Program` knows the `Linkage` it belongs to, and allows for a `TargetProgram` to be looked up based on a specific `TargetRequest`. A `TargetProgram` then allows layout information and compiled kernel code to be asked for on-demand, in order to support eventual "live" compilation scenarios.
* The `EndToEndCompileRequest` type is a composition/coordination type that replaces the old `CompileRequest` in a way that uses the services of the various other types. It owns a few pieces of state that only make sense in the context of an end-to-end compile (e.g., there is really no way to "pass through" code when the front- and back-ends are run separately) or a command-line compile (everything to do with specifying output paths for files is really just for the benefit of `slangc`, and might even be moved there over time).
* One important detail is that the `EndToEndCompilRequest` owns all of the string-based generic arguments for both global and entry-point generic parameters. The logic in `check.cpp` for dealing with those arguments has been heavily refactored to separate out the parsings steps that are specific to end-to-end compilation with string-based type arguments, and the semantic checking steps that result in a specialized `Program` (which can be exposed through new APIs that aren't tied to end-to-end compilation).
It is perhaps not surprising that this change had a lot of consequences, so I'll briefly run over some of the main categories of changes required:
* I changed the way that global generic arguments are passed via API (use `spSetGlobalGenericArgs` instead of the generic arguments for `spAddEntryPointEx`, which are not just for entry-point generics), which has been a change that we've needed for a long time. This is technically a breaking API change, although we should have very few client applications that care about it.
* A bunch of places that used to take "big" objects like `CompileRequest` now just take the sub-pieces they care about (e.g., a function might have only needed a `Linkage` and a `DiagnosticSink`). This makes many subroutines or "context" struct types more generally useful, at the cost of taking more parameters.
* In a few cases the conceptually clean separation of the layers breaks down (often for edge-case or compatibility features), and so we may pass along additional objects that are allowed to be null, but are used when present. A big example of this is how the back-end code generation routines accept an `EndToEndCompileRequest` that is optional, and only used to check whether "pass through" compilation is needed. We should probably look into cleaning this kind of logic up over time so that we don't need to violate the apparent separation of phases of compilation.
* In cases where separation of layers was being broken for the sake of GLSL features, I went ahead and ripped them out, since all of that should be dead code anyway.
* In many cases I increased the encapsulation of data in the core types to help track down use sites and make sure they are following invariants better.
* In cases where code was doing, e.g., `context->shared->compileRequest->session->getThing()` I have tried to introduce convenience routines so that the usage site is just `context->getThing()` to improve encapsulation and allow changes to be made more easily going forward.
* The `noteInternalErrorLoc` functionality was moved off of the compile request and into `DiagnosticSink`, since that is the one type you can rely on having around when you want to note an internal error. We may consider going forward if (and how) it should reset the counter used for noting locations on internal errors.
* A few APIs now take `DiagnosticSink*` arguments where they didn't before, and as a result some public APIs need to create `DiagnosticSink`s to pass in, before going ahead and ignoring the messages. In the future there should be variations of these APIs that accept an `ISlangBlob**` parameter for the output.
* fixup: missing include for compilers with accurate template checking (non-VS)
* fixup: review feedback
|
|
* Initial support for dynamic dispatch using "tagged union" types
Suppose a user declares some generic shader code, like the following:
```hlsl
interface IFrobnicator { ... }
type_param T : IFrobincator;
ParameterBlock<T : IFrobnicator> gFrobnicator;
...
gFrobincator.frobnicate(value);
```
and then they have some concrete implementations of the required interface:
```hlsl
struct A : IFrobnicator { ... }
struct B : IFrobnicator { ... }
```
The current Slang compiler allows them to generate distinct compiled kernels for the case of `T=A` and the case of `T=B`. This means that the decision of which implementation to use must be made at or before the time when a shader gets bound in the application.
This change adds a new ability where the Slang compiler can generate code to handle the case where `T` might be *either* `A` or `B`, and which case it is will be determined dynamically at runtime. This means a single compiled kernel can handle both cases, and the decision about which code path to run can be made any time before the shader executes.
This new option is supported by defining a *tagged union* type. Via the API, the user specifies that `T` should be specialized to `__TaggedUnion(A,B)` (the double underscore indicates that this is an experimental and unsupported feature at present). We refer to the types `A` and `B` here as the "case" types of the tagged union. Conceptually, the compiler synthesizes a type something like:
```hlsl
struct TU { union { A a; B b; } payload; uint tag; }
```
The user can then allocate a constant buffer to hold their tagged union type, and when they pick a concrete type to use (say `B`), they fill in the first `sizeof(B)` bytes of their buffer with data describing a `B` instance, and then set the `tag` field to the appopriate 0-based index of the case type they chose (in this case the `B` case gets the tag value `1`).
Actually implementing tagged unions takes a few main steps:
* Type parsing was extended to special-case `__TaggedUnion` as a contextual keyword. This is really only intended to be used when parsing types from the API or command-line, and Bad Things are likely to happen if a user ever puts it directly in their code. Eventually construction of tagged unions should be an API feature and not part of the language syntax.
* Semantic checking was extended to recognize that a tagged union like `__TaggedUnion(A,B)` shoud support an interface like `IFrobnicator` whenever all of the case types suport it, as long as the interface is "safe" for use with tagged unions (which means it doesn't use a few of the advancd langauge features like associated types).
* The IR was extended with instructions to represent tagged union types and to extract their tag and the payload for the different cases as needed.
* IR generation was extended to synthesize implementations of interface methods for any interface that a tagged union needs to support. Right now the implementation is simplistic and only handles simple method requirements, which it does by emitting a `switch` instruction to pick between the different cases.
* A new IR pass was introduced to "desugar" any tagged union types used in the code. The downstream HLSL and GLSL compilers don't support `union`s, so we have to instead emit a tagged union as a "bag of bits" and implement loading the data for particular cases from it manually.
* Final code emit mostly Just Works after the above steps, but we had to introduce an explicit IR instruction for bit-casting to handle the output of the desugaring pass.
There are a bunch of gaps and caveats in this implementation, but that seems reasonable for something that is an experimental feature. The various `TODO` comments and assertion failures in unimplemented cases are intended, so that this work can be checked in even if it isn't feature-complete.
* fixup: missing files
* fixup: typos
|
|
Fixes #775
It was reported (in #775) that Slang doesn't handle initializer-list syntax when initializing matrix variables. When starting on a fix for that it became apparent that the time was right to fix two broad issues in the compiler's current handling of `{}`-enclosed initializer lists.
The first issue was that the front-end checking of initializer lists wasn't handling the C-style behavior where an initializer list can either contain nested `{}`-enclosed lists for sub-arrays/-structures, or directly contain "leaf" values for initializing those aggregates. For example, the following two variable declarations ought to be equivalent:
```hlsl
int4 a[] = { {1, 2, 3, 4}, {5, 6, 7, 8} };
int4 b[] = { 1, 2, 3, 4, 5, 6, 7, 8 };
```
Getting this distinction right is important because we want to support initializing a matrix either from a list of vectors for its rows, or a list of scalars for its elements (in row-major order).
The front-end semantic checking logic for initializer lists was revamped so that it conceptually tries to "read" an expression of a desired type from the initializer list, and decides at each step whether to consume a single expression by coercing it to the desired type, or to recursively read multiple sub-values to construct the type as an aggregate. The logic for deciding between direct vs aggregate initialization could potentially use some tweaking, but luckily it should always handle the case where users introduce explicit `{}`-enclosed sub-lists to make their intention clear, so that existing Slang code should continue to work as before.
The second issue was that initializers without the expected number of elements weren't implemented in code generation, so they would lead to internal compiler errors. This change revamps the codegen logic for initializer lists so that it can synthesize default values for fields/elements that were left out during initialization. This includes an attempt to support default initialization of `struct` fields based on explicitly written initialization expressions.
|
|
* First step toward supporting use of interfaces as existential types
Traditional generics involve universal quantification. E.g., a declaration like:
```
void drive<T : IVehicle>(T vehicle);
```
indicates for *for all* types `T` that implement the `IVehicle` interface, the `drive()` function is available.
In contrast, whend directly using an interface type like:
```
IVehicle v = ...;
v.doSomething();
```
we only know that there *exists* some concrete type (we could call it `E`) such that `v` refers to a value of type `E`, and `E` implements the `IVehicle` interface. In order to perform an operation like `v.doSomething()` we need to "open" the existential value so that we can look at the concrete type and how it implements the `IVehicle.doSomething` requirement.
This change adds a very explicit representation of existentials to Slang's IR. An operation like `e = makeExistential(v, w)` creates a value of some existential type (interfaces being our only existential types for now), by wrapping a concrete value `v` (the type of `v` can be seen as an implicit operand) and a witness table `w` showing that the type of `v` implements the requirements of the chosen interface type.
In turn, opening of an existential is handled with operations `extractExistential{Value|Type|WitnessTable}` which pull the corresponding piece of information out of a value of existential type (which somewhere in the code had to have been created with `makeExistential`).
The change includes a trivial simplification pass that can detect cases where an `extractExistential*` operation is applied direclty to a `makeExistential` operation, so that there is only one possible result that could be extracted. This allows for simplification of existential types used in trivial ways for local variables (this is mostly so I can check in a functional test, rather than to actually support useful code involving interfaces right now).
The logic in the semantic checking phase of the compiler is comparatively more complex.
When we are about to perform member lookup given an expression like `obj.member` we will first check if `obj` has an existential type, and if it does we will construct a suitable local context in which we extract the value, type, and witness table from the existential (these all become explicit AST expression nodes), and then use the extracted value as the base of the lookup operation.
The nature of existential values is that two different values with the same existential (interface) type could wrap concrete values with differnt types, so that we need to carefully refer only to the extracted type/value/witness-table of specific *values*. We handle this right now by conceptually moving the existential-type value into a local variable (by introducing a `LetExpr` that amounts to `let v = <init> in <body>`) and then require that the extract expressions must refer to the (immutable) variable declaration from which they are extracting a value.
(Eventually we should expand this so that when using an immutable local variable of existential type we just use that variable as-is rather than introduce a new temporary)
A simple test case is included that uses an interface type in an almost trivial way for a local variable; this test can be run and produces the expected results.
A more complex test case that passes an existential into a function is included, but left disabled because a more aggressive simplification approach is required to generate working code from it.
* Add missing file for expected test output
* Fixups for merge from top-of-tree
|
|
* Specialize away resource-type function parameters
Work on #397.
Introduction
------------
Suppose a user writes a function that takes a resource type as a parameter:
```hlsl
float4 getThing(RWStructuredBuffer<float4> buffer, int index)
{
return buffer[index];
}
```
This function creates challenges when generating code for GLSL-based targets, because a global shader parameter of type `RWStructuredBuffer`:
```hlsl
RWStructuredBuffer<float4> gBuffer;
```
translates to a global GLSL `buffer` declaration:
```hlsl
buffer _S0
{
float4 _data[];
} gBuffer;
```
There is no equivalent to that `buffer` declaration that can be used in function parameter position, and it is illegal in GLSL to pass `gBuffer` into a function.
(Aside: yes, we could in principle translate a function parameter like `RWStructuredBuffer<float4> buffer` to `float4 buffer[]`, but that will not in turn generalize to arrays of structured buffers; it is a dead-end strategy)
The solution employed by many shader compilers is to "inline everything" to eliminate the need for parameters of resource types, and then rely on dataflow optimization to eliminate locals of resource types. This strategy can of course lead to an increase in code size, and it also means that call stacks are lost when doing step-through debugging. Another serious issue is that an "early `return`" from a function can turn into the equivalent of a multi-level `break` when inlined, and not all of our targets support multi-level `break`.
The solution implemented in this change works around some, but not all, of the problems with full inlining.
The approach here generates specialized versions of a function like `getThing`, adapted to the actual arguments provided at different call sites.
Thus if we have code like:
```hlsl
RWStructuredBuffer<float4> gA;
RWStructuredBuffer<float4> gB[10];
...
getThing(gA, x);
getThing(gA, y);
getThing(gB[someVal], z);
```
we will generate two specializations of `getThing`: one specialized for the `buffer` parameter being `gA` and the other for `gB`:
```hlsl
float4 getThing_gA(int index) { return gA[index]; }
float4 getThing_gB(int _val, int index) { return gB[_val][index]; }
```
and the call sites will change to match:
```hlsl
getThing_gA(x);
getThing_gA(y);
getThing_gB(someVal, z);
```
Note how in the case where the argument being passed in was obtained by indexing into an array of resources, the callee is specialized to the identity of the global shader parameter (`gB`), and now accepts a new parameter to indicate the array index into it.
While this description motivates the change based on GLSL output, the same basic issue can arise for other targets.
For example, while current HLSL has added the `ConstantBuffer<T>` type, it is not supported on older targets, and it turns out that even dxc does not allow functions to have `ConstantBuffer<T>` parameters.
Longer-term, we will likely need to do even more aggressive specialization both in order to generate SPIR-V output directly, and also to deal with function that have return values or `out` parameters of resource types.
Implementation
--------------
The meat of the change is in `ir-specialize-resources.{h,cpp}`, where we have a pass that looks at all call sites (`IRCall` instructions) in the program, and attempts to replace them with calls to specialized functions, where the specializations are generated on-demand.
The code in this pass is heavily commented, so hopefully it serves to explain itself all right.
After specialization is complete, we may still have functions like the original `getThing` that will produce invalid code when emitted as GLSL, so we need a way to make sure they don't appear in the output.
To date we've had some very ad hoc approaches for ignoring IR constructs that we don't want to affect emitted code, but this change goes ahead and adds a more real dead code elimination (DCE) pass in `ir-dce.{h,cpp}`.
This pass follows a straightforward approach of tagging instructions that are "live" and then propagating liveness through the whole program, before making a single pass to delete anything that isn't live.
When I first added the DCE pass it eliminated *everything* because there were no "roots" for liveness.
I solved this for now by adding a new decoration, `IREntryPointDecoration`, to mark shader entry points in the IR which should always be live (as should anything they depend on).
A secondary problem that arose was that for GLSL ray tracing shaders it is possible for the incoming/outgoing payload or attributes parameters to be unused, but eliminating them as dead would change the signature of a shader an potential break the rules for how ray tracing programs communicate.
I added a very simple `IRDependsOnDecoration` that allows one IR instruction to keep another alive *as if* it used it, without actually using it.
There's also a fixup in the IR dumping logic where I was forgetting to store anything in the mapping from instruction to their names, so that the name of an instruction was getting incremented each time it was referenced.
Testing
-------
There are three different tests added as part of this change:
* The `compute/func-resource-param` test covers the basic `RWStructuredBuffer` case above, which we expect to work fine for D3D11/12, but fail for Vulkan without specialization.
* The `cross-compile/func-resource-param-array` test covers the case where we don't just have one resource, but an array of them. This is not an end-to-end compute test primarily because our `render-test` application doesn't yet handle arrays of resources correctly in its binding logic.
* The `compute/func-cbuffer-param` test covers the case of a function with a `ConstantBuffer<T>` parameter, which requires specialization to become valid for any of our targets.
* fixup: warnings/errors from other compilers
* fixup: typos and cleanup
* fixup: typos
|
|
Before this change, global shader parameters were represented in the IR as just being ordinary global variables.
The only indication that a particular global represented a parameter was when it got a layotu attached to it (as part of back-end processing), and we've had a number of bugs related to layouts being dropped so that what should have been a shader parameter turned into an ordinary global variable in the output.
This change is more strongly motivated by the fact that making shader parameters look like globals means that we cannot easily reason about their value when doing IR transformations.
If we see two `load`s from the same global variable can we assume they yield the same value?
In the general case we cannot, and this means that any transformation that wants to rely on the fact that an input `Texture2D` shader parameter can't actually change over the life of the program needs to do extra work.
The fix here is to introduce a new kind of IR instruction that represents a global shader parameter directly (not a pointer to it as a global would), at which point there isn't even such a notion as a "load" from the parameter, since it represents the value directly.
In several cases logic that used to apply to global variables in case they were shader parameters (by looking for a layout) is now moved to apply to these global parameters.
The biggest source of issues in this change was that switching from pointers to plain values to represent these shader parameters stresses different cases in type legalization. I also had to deal with the case of legalization for GLSL where we actually *do* need global shader parameters that are writable (since varying output goes in the global scope), but in that case I borrowed the use of pointer-like `Out<...>` and `InOut<...>` types to represent that intent, which we were already using for function parameters representing outputs.
A few tests started failing because the changes lead to a slightly different order of code emission, which in some HLSL tests resulted in a function parameter named `s` getting emitted before a global parameter named `s`, leading to the latter getting the name `s_1` instead of `s_0`.
A few SPIR-V tests started failing because the new approach means that we no longer end up performing a load from all varying input parameters at the start of `main` and instead reference the varying inputs directly. The resulting code is more idomatic, but it differed from the baselines for those tests.
|
|
* Move mangled name out of IRGlobalValue
Previously the `IRGlobalValue` type was used as a root for all IR instructions that can have "linkage," in the sense that a definition in one module can satisfy a use in another module.
The mangled symbol name was stored in state directly on each `IRGlobalValue`, which created some complications, and also forced IR instructions that wanted to support linkage to wedge into the hierarchy at that specific point.
This change moves the mangled name out into a decoration: either an `IRImportDecoration` or an `IRExportDecoration`, both of which inherit from `IRLinkageDecoration` which exposes the mangled name.
This change has a few benefits:
* We can now have any kind of instruction be exported/imported, without having to inherit from `IRGlobalValue`. This could potentially let `IRStructType` and `IRWitnessTable` be simplified to just have operand lists instead of dummy chldren as they do today.
* We can now easily have "global values" like functions that explicitly *don't* get linkage, instead of using a null or empty mangled name as a marker.
* We can use the exact opcode on a linkage decoration to distinguish imports from exports, which could be used to more accurately resolve symbols during the linking step.
Other than adding the decorations and making sure that AST->IR lowering adds them, the main changes here are around any code that used `IRGlobalValue`. Variables and parameters of type `IRGlobalValue*` were changed to `IRInst*` easily, so the main challenge was around code that *casts* to `IRGlobalValue*.
In cases where a cast to `IRGlobalValue` also performed a test for the mangled name being non-null/non-empty, we simply switched the code to check for the presence of an `IRLinkageDecoration`, since that is the new way of indicating a value with linakge.
Most of the serious complications arose in `ir.cpp` around the "linking"/target-specialization and generic specialization steps.
The "linking" logic was checking for `IRGlobalValue` to opt into some more complicated cloning logic, and just checking for a linkage decoration here wasn't sufficient since the front-end *does* produce global values without linkage in some cases (e.g., for a function-`static` variable we produce a global variable without linkage). This logic was updated to just check for the cases that used to amount to `IRGlobalValue`s directly by opcode. It might be simpler in the short term to have kept `IRGlobalValue` around to make the existing casts Just Work, but I'm confident that this logic could actually be rewritten for much greater clarity and simplicity and that is the better way forward.
The generic specialization logic was using some really messy code to generate a new mangled name to represent the specialized symbol, and then searching for an existing match for that name.
The original idea there was that an IR module could include "pre-specialized" versions of certain generics to speed up back-end compilation by eliminating the need to specialize in some cases, but this feature has never been implemented so the overhead here is just a waste.
Instead, I moved generic specialization to use a simpler dictionary to map the operands to a `specialize` instruction over to the resulting specialized value.
This allows for some simplifications in the name mangling logic, because it no longer needs to figure out how to produce mangled names from IR instructions representing types/values.
As part of this change I also overhauled the IR emit logic to produce cleaner output by default, borrowing some of the ideas from the logic in `emit.cpp`. IR values are now automatically given names based on their "name hint" decoration, if any, to make the code easier to follow, and I also made it so that types and literals get collapsed into their use sites in a new "simplified" IR dump mode (which is currently the default, with no way to opt into the other mode without tweaking the code). The resulting IR dumps are much nicer to look at, but as a result the one test that involves IR dumping (`ir/string-literal`) doesn't really test what it used to.
One weird issue that came up during testing is that the `transitive-interface` test had previously been producing output that made no sense (that is, the expected output file wasn't really sensible), and somehow these changes were altering its behavior. Changing the test to use `int` values instead of `float` was enough to make the output be what I'd expect, and hand inspection of generating DXBC has me convinced we were compiling the `float` case correctly too. There appears to be some issue around tests with floating-point outputs that we should investigate.
* fixup: C++ declaration order
|
|
* Make a test case use IR serialization
* Make all IR instructions usable as parents
This makes it so that every `IRInst` has the list of children that used to be on `IRParentInst` and eliminates `IRParentInst`.
Most places in the code were only checking against `IRParentInst` so that they could know whether there were child instructions to iterate over.
This change bloats the size of every instruction by two pointers, but we hope to be able to eliminate that overhead with a better encoding later.
* Change IR decorations to be instructions.
The main change here is that `IRDecoration` now inherits from `IRInst`, and `IRInst` now has a single linked list that holds both decorations *and* children.
At each point where code used to loop over `getChildren()` on an `IRInst`, I checked whether it made sense to leave the operation as processing just the children, or if it should process both decorations and children.
The thorniest bit was making sure the logic for inserting an instruction into a parent is correct. For the most part, once IR code is built all insertions are explicitly before/after another instruction, so the ordering can't get messed up. The sticking point is any code that does an explicit `insertAtStart` or `insertAtEnd`, but I surveyed those to make sure they are correct in context, and I also made all insertions bottleneck through one routine that does a better job of asserting the preconditions than what was there before. We may still want a "smart" insertion function at some point so that if somebody does `someDecoration->insertAtEnd(someInst)` the decoration intelligently goes to the end of the decoration list, and not the entire decorations-and-children list.
All of the existing decoration types were refactored to provide accessors for their operands, rather than directly exposing fields. In most cases the operands are required to be `IRConstant` nodes of fixed types. Not all of these types need to be kept around in the new approach, but they were left in so that as much existing code as possible can be kept working.
The `IRBuilder` was extended with factory functions to make the various decoration types and attach them.
All the fields in concrete decorations that were using `StringRepresentation` or `Name` pointers are now using IR-level string operands which provide their value as an `UnownedStringSlice`, so logic that was working with those decoration values needed to be updated here and there. I also needed to add the logic to clone string-literal values to the IR cloning pass, since they are now being used in almost every piece of code.
A new type of constant IR instruction for literal pointers was added, to handle the cases where an IR decoration needs an operand that is a raw AST-level pointer. These are even being serialized, although we obviously should not rely on them to round-trip through serialization in the future. Ideally, a follow-on change should add a cleanup pass where we remove any decorations from a module that shouldn't be allowed in the serialized code.
The biggest overall cleanup is in the serialization logic, where a lot of code just disappears because it can process the raw "decorations and children" list as the logical children of an IR instruction. The only special cases left are literals (which seem like they will always need special-casing) and global values (because they have a mangled name, which we plan to move into a decoration).
One other example of a simplification made possible by this change: the `IRNotePatchConstantFunc` instruction was implemented as an instruction only because it couldn't be encoded as a decoration at the time (it needed to have an operand that referenced an IR function).
The IR dumping logic was also updated (which meant a change to the `ir/string-literal` test) to try to make it print out all decorations a bit more systematically now that they are encoded like other instructions. The formatting isn't quite perfect, but it is good enough to be able to read what is going on.
I didn't include updates to the validation logic to ensure that decorations are being added in ways that follow the invariants, but that would be a nice thing to add next.
* fixup: 64-bit issues
* fixup: forward declaration issues
|
|
* Change how buffers are emitted
This is a change with a lot of pieces, which can't always be separated out cleanly. I'm going to walk through them in what I hope is a logical order.
The main goal of this change was to allow arrays of structured buffers to translate to Vulkan. Consider two declarations of structured buffers in HLSL/Slang:
```hlsl
StructuredBuffer<X> single;
StructuredBuffer<Y> multiple[10];
```
The current translation logic was handling `single` by translating it into an *unnamed* GLSL `buffer` block like:
```glsl
layout(std430)
buffer _S1
{
X single[];
};
```
That syntax allows an expression like `single[i]` in Slang to be translated simply as `single[i]` in GLSL.
But that naive translating doesn't work for `multiple`, since we need to declare a array of blocks in GLSL, which requires giving the whole thing a name:
```glsl
layout(std430)
buffer _S2
{
Y _data[];
} multiple[10];
```
Now a reference to `multiple[i][j]` in Slang needs to become `multiple[i]._data[j]` in GLSL.
To avoid having way too many special cases around single structured buffers vs. arrays, it makes sense to allows emit things in the latter form, so that we instead lower `single` as:
```glsl
layout(std430)
buffer _S1
{
X _data[];
} single;
```
So that now a reference to `single[i]` becomes `single._data[i]` in GLSL.
Most of that can be handled in the standard library translation of the structured buffer indexing operations.
The only wrinkle there is that there were some *old* special-case instructions in the IR intended to handle buffer load/store operations (these were added back when I was trying to keep the "VM" path working). These aren't really needed to have structured-buffer operations work; they can be handled as ordinary functions as far as the stdlib is concerned. I removed the old instructions.
Along the way, it became clear that a few other cases follow the same pattern. Byte-addressed buffers are an obvious case. We were lowering HLSL/Slang:
```hlsl
ByteAddressBuffer b;
...
uint x = b.Load(0);
```
to GLSL like:
```glsl
layout(std430)
buffer _S1
{
uint b[];
};
...
uint x = b[0];
```
That logic would fail for arrays the same way that the structured buffer case was failing. The fix is the same: use named `buffer` blocks and then introduce an explicit `_data` field:
```glsl
layout(std430)
buffer _S1
{
uint _data[];
} b;
...
uint x = b._data[0];
```
Just like with structured buffers, all of the VK translation for operations on byte-addressed buffers can be implemented directly in teh stdlib, so once the emit logic was changed it was just a matter of adding `._data` to a bunch of VK tranlsations.
It turns out that arrays of constant buffers have more or less the same problem, and furthermore we have some problems with any code that directly uses the modern HLSL `ConstantBuffer<T>` type.
Note: the emit logic around constant buffers sometimes refers to "parameter groups" because that is being used in the compiler as a catch-all term for constant buffers, texture buffers, and parameter blocks.
The existing code was going out of its way to reproduce the way that constant buffer declarations are implicitly referenced in HLSL:
```hlsl
cbuffer C { float f; }
...
float tmp = f; // No reference to `C` here
```
This can be seen in the emit logic with the `isDerefBaseImplicit` function, which is used to take the internal IR representation for a reference to `f` (which is closer to the expression `(*C).f` or `C->f`) and leave off any reference to `C` so that we emit just `f`.
That kind of logic just flat out doesn't work in some important cases. Arrays of constant buffers are a clear one:
```hlsl
ConstantBuffer<X> cbArray[3];
...
X x = cbArray[0];
```
There is no way to translate that to an ordinary `cbuffer` declaration at all. The same problem can be created without arrays, though:
```hlsl
ConstantBuffer<X> singleCB;
...
X x = singleCB;
```
The current strategy for translating constant buffers was translating `singleCB` into a `cbuffer` declaration that reproduced the fields of `X` as its members, which just wouldn't work:
```hlsl
cbuffer singleCB
{
float f; // field of `X`
}
...
X x = singleCB; // ERROR: there is nothing named `singleCB` in this HLSL
```
The new strategy is more consistent. We still generate a `cbuffer` declaration for a single constant buffer, but we always give it a single field of the chosen element type:
```hlsl
cbuffer singleCB
{
X singleCB;
}
...
X x = singleCB; // this works fine!
```
And in the array case we generate code that uses the explicit `ConstantBuffer<T>` type:
```hlsl
ConstantBuffer<X> cbArray[3];
...
X x = cbArray[0];
```
The GLSL output is more complicated because unlike with HLSL there is no implicit conversion from a uniform block to its element type (there is no notion of an element type). The array case thus needs a `_data` field similar to what we do for structured buffers:
```glsl
layout(std140)
uniform _S3
{
X _data;
} cbArray[3];
...
X x = cbArray[0]._data;
```
And then the non-array case needs to have a similar `_data` field for consistency:
```glsl
layout(std140)
uniform _S1
{
X _data;
} singleCB;
...
X x = singleCB._data;
```
This is handled by inserting the necessary reference to `_data` whenever we dereference a constant buffer, either as part of a load instruction (loading from the whole CB as a pointer), or an `IRFieldAddress` instruction which forms a pointer into the CB (e.g., `&(singleCB->f)` becomes `singleCB._data.f`).
The current emit logic handles `ParameterBlock<X>` differently from `ConstantBuffer<X>`, but really only to allow parameter blocks to be explicitly named in the output, while constant buffers were left implicit by default. Thus the only difference was a legacy one (from back when trying to exactly reproduce the HLSL text we got as input was considered an important goal), and the new approach to emitting constant buffers would get rid of it.
I removed the separate logic for emitting `ParameterBlock<X>` and just let the handling for constant buffers deal with it.
Note that any resource types inside of a `ParameterBlock<X>` would have been moved out as part of legalization, so that a parameter block is 100% equivalent to a constant buffer when it comes time to emit code.
Unsurprisingly, changing the way we generate HLSL and GLSL output for all these buffer types meant that any tests that were directly comparing the output of `slangc` against `fxc`, `dxc`, or `glslang` broke.
The basic approach to fixing the breakage in GLSL tests was to update the GLSL baseline to reflect the new output startegy. In some cases I used macros to name the various `_S<digits>` temporaries so that future renaming will hopefully be easier (it would be great if we auto-generated temporary names with a bit more context). There was one GLSL test (`tests/bugs/vk-structured-buffer-binding`) that was using raw GLSL expected output, and this was changed to use a GLSL baseline to generate SPIR-V for comparison.
For HLSL tests we were sometimes running the same input file through `slangc` and `fxc`/`dxc`, and in these cases I macro-ized the various `cbuffer` declarations to generate different declarations depending on the compiler.
I completely dropped the tests coming from the D3D SDK because they aren't providing much coverage, and updating them would change them so far from the original code that the purported benefit (using a body of existing shaders) would be lost.
I also dropped the explicit matrix layout qualifiers in the `matrix-layout` test because the new output strategy breaks those for GLSL (you can't put matrix layout qualifiers on `struct` fields, and now the body of every constant buffer is inside a `struct`). This isn't as big of a loss as it seems, because our handling of those qualifiers wasn't really right to begin with. Slang users should only be setting the matrix layout mode globally (and we should probably switch to error out on the explicit qualifiers for now).
The other thing that got dropped is tests involving `packoffset` modifiers.
Slang already warns that it doesn't support these, and the way they were used in the test cases is actually misleading. For the binding/layout-related tests, the goal was to show that Slang reproduces the same layout as fxc, in which case explicitly enforcing a layout via `packoffset` seems like cheating (are we sure we enforced the layout fxc would have produced?). The real reason was that Slang used to emit explicit `packoffset` on *every* field of a `cbuffer` it would output, because of an `fxc` bug where you couldn't use `register` on textures/samplers declared inside a `cbuffer` unless *every* field in the `cbuffer` used a `register` or `packoffset` modifier. Slang hasn't required that behavior in a while because it now splits textures and samplers, and the one test case where we needed `packoffset` to work around the `fxc` bug in the baseline HLSL has been macro-ified even more to work around the bug.
The amount of churn in the test cases is unfortunate, but it continues to point at the weakness of any testing strategy that checks for exact equivalent between Slang's output and that of other compilers. We need to keep working to replace these tests with better alternatives.
In `check.cpp` there is logic to perform implicit dereferencing, so that if you write `obj.f` where `obj` is a `ConstantBuffer<X>` (or some other "pointer-like" type) and `f` is a field in `X`, then this effectively translates as `(*obj).f`. That is, we dereference the value of type `ConstantBuffer<X>` to get a value of type `X`, and then refer to the field of the `X` value.
There was a problem where the logic to insert that kind of implicit dereference operation was using a reference (`auto& type = ...`) for the type of the expression being dereferenced, and then clobbering it. This would mean that an expression of type `ConstantBuffer<X>` would have its type overwritten to be just `X` and then codegen would break later on.
I'm not sure how we haven't run into that before.
The `array-of-buffers` test case was added to confirm that we now support arrays of constant, structured, and byte-address buffers for both DXIL and SPIR-V output.
Okay, so that was a lot of stuff, but hopefully it is clear how this all works to make the output of the compiler more consistent and explicit, while also supporting the required new functionality.
* fixup: review feedback
|
|
* Add Vulkan cross-compilation for byte-address buffers
This covers `ByteAddressBuffer`, `RWByteAddressBuffer`, and `RasterizerOrderedByteAddressBuffer`. A declaration of any of these types translates to a GLSL `buffer` declaration with a single `uint` array of data. Most of the methods on these types then have straightforward translations to operations on the array. The overall translation is similar to what was already being done for structured buffers.
While implementing GLSL translation for the various atomic (`Interlocked*`) methods, I discovered that some of these included declarations that aren't actually included in HLSL. I cleaned these up, including in the declarations of the global `Interlocked*` functions.
The test case that is included here covers only the most basic functionality: `Load`, `Load2`, `Load4` and `Store`. We should try to back-fill tests for the remaining methods when we have time.
Two large caveats with this work:
1. We don't handle arrays of byte-address buffers, just as we don't handle arrays of structured buffers. That will take additional work.
2. We don't handle byte-address (or structured) buffers being passed as function parameters, since the parameter would need to be declared as a bare `uint[]` array.
* Fixup: don't lump raytracing acceleration structures in with buffers
Raytracing acceleration structures share a common base class with byte-address buffers (they are both buffer resources without a specific element type), and I was mistakently matching on this base class in an attempt to have a catch-all that applied to all byte-address buffers.
The fix here was to add a distinct base class for all byte-address buffers and catch that instead.
* Fixup: typos
|
|
* Add a warning on missing return, and initial SCCP pass
The user-visible feature added here is a diagnostic for functions with non-`void` return type where control flow might fall off the end. This *sounds* like a trivial diagnostic to add as part of the front-end AST checking, but that can run afoul of really basic stuff like:
```hlsl
int thisFunctionisOkay(int a)
{
while(true)
{
if(a > 10) return a;
a = a*2 + 1;
}
// no return here!
}
```
This function "obviously" doesn't need to have a `return` statement at the end there, but realizing this fact relies on the compiler to understand that the `while(true)` loop can't exit normally, and doesn't contain any `break` statement. One can write "obvious" examples that need more and more complex analysis to rule out.
The answer Slang uses for stuff like this is to do the analysis at the IR level right after initial code generation (this would be before serialization, BTW, so that attached `IRHighLevelDeclDecoration`s can be used).
When lowering the AST to the IR, we always emit a `missingReturn` instruction (a subtype of `IRUnreachable`) at the end of its body if it isn't already terminated. The IR analysis pass to detect missing `return` statements is then as simple as just walking through all the functions in the module and making sure they don't contain `missingReturn` instructions.
For that simple pass to work, we first need to make some effort to remove dead blocks that control flow can never reach. This change adds a very basic initial implementation of Spare Conditional Constant Propagation (SCCP), which is a well-known SSA optimization that combines constant propagation over SSA form with dead code elimination over a CFG to achieve optimizations that are not possible with either optimization along.
For the moment, we don't actually implement any constant *folding* as part of the SCCP pass, so we can eliminate the dead block in a case like the function above (and those in the test case added in this change), but will not catch things like a `while(0 < 1)` loop. Handling more "obvious" cases like that is left for future work.
* fixup: warning on unreachable code
* Handle case where user of an inst isn't in same function/code
The code as assuming any instruction in the SSA work list has to come from the function/code being processed, but this misses the case where an instruction in a generic has a use inside the function that the generic produces.
This change adds code to guard against that case.
|
|
* Move to newer glslang
* Support cross-compilation of ray tracing shaders to Vulkan
This change allows HLSL shaders authored for DirectX Raytracing (DXR) to be cross-compiled to run with the experimental `GL_NVX_raytracing` extension (aka "VKRay").
* The GLSL extension spec is marked as experimental, so that any shaders written using this support should be ready for breaking changes when the spec is finalized.
* "Callable shaders" are not exposed throug the GLSL extension, so this feature of DXR will not be cross-compiled.
* The experimental Vulkan raytracing extension does not have an equivalent to DXR's "local root signature" concept. This does not visibly impact shader translation (because the local/global root signature mapping is handled outside of the HLSL code), but in practice it means that applications which rely on local root signatures on their DXR path will not be able to use the translation in this change as-is; more work will be needed.
The simplest part of the implementation was to go into the Slang standard library and start adding GLSL translations for the various DXR operations.
In some cases, like mapping `IgnoreHit()` to `ignoreIntersectionNVX()` this is almost trivial.
The various functions to query system-provided values (e.g., `RayTMin()`) were also easy, with the only gotcha being that they map to variables rather than function calls in GLSL, and our handling of `__target_intrinsic` assumes that a bare identifier represents a replacement function name, and not a full expression, so we have to wrap these definitions in parentheses.
The tricky operations are then `TraceRay<P>()` and `ReportHit<A>()`, because these two are generics/templates in HLSL.
GLSL doesn't support generics, even for "standard library" functions, so the raytracing extension implements a slightly complex workaround: the matching operations `traceNVX()` and `reportIntersectionNVX()` pass the payload/attributes argument data via a global variable.
That is, shader code for the GLSL extensions writes to the global variable and then calls the intrinsic function.
The linkage between the call site and the global is established by a modifier keyword (`rayPayloadNVX` and `hitAttributeNVX`, respectively) and in the case of ray payload also uses `location` number to identify which payload global to use (since a single shader can trace rays with multiple payload types).
Our translation strategy in Slang tries to leverage standard language mechanisms instead of special-case logic.
For example, to translate the `ReportHit<A>()` function, we provide both a default declaration that will work for HLSL (where the operation is built-in with the signature given), and a *definition* marked with the `__specialized_for_target(glsl)` modifier.
The GLSL definition declares a function `static` variable that will fill the role of the required global, and then does what the GLSL spec requires: assigns to the global, and then calls the `reportIntersectionNVX` builtin (which we declare as a separate builtin).
Our ordinary lowering process will turn that `static` variable into an ordinary global in the IR, and the `[__vulkanHitAttributes]` attribute on the variable will be emitted as `hitAttributeNVX` in the output.
There is no additional cross-compilation logic in Slang specific to `ReportHit<A>()` - the target-specific definition in the standard library Just Works.
The case for `TraceRay<P>()` is a bit more complicated, simply because the GLSL `traceNVX()` function needs to be passed the `location` for the payload global.
We implement the payload global as a function-`static` variable, with the knowledge that every unique specialization of `TraceRay<P>()` will generate a unique global variable of type `P` to implement our function-`static` variable.
We then add a slightly magical builtin function `__rayPayloadLocation()` that can map such a variable to its generated `location`; the logic for this is implemented in `emit.cpp` and described below.
We also changed the `RayDesc` and `BuiltinTriangleIntersectionAttributes` types from "magic" intrinsic types over to ordinary types (because the GLSL output needs to declare them as ordinary `struct` types).
This ends up removing some cases in the AST and IR type representations.
By itself this change would break HLSL emit, because in that case the types really are intrinsic.
We added a `__target_intrinsic` modifier to these types to make them intrinsic for HLSL, and then updated the downstream passes to handle the notion of target-intrinsic types.
The logic for binding/layout of entry point inputs and outputs was updated so that raytracing stages don't follow the default logic for varying input/output parameters.
This is because the input/output parameters of a raytracing entry point aren't really "varying" in the same sense as those in the rasterization pipeline.
In particular, the SPIR-V model for raytracing input and output treats "ray payload" and "hit attributes" parameters as being in a distinct storage class from `in` or `out` parameters.
We also detect cases where a ray tracing stage declares inputs/outputs that it shouldn't have. This logic could conceivably be extended to other stages (e.g., to give an error on a compute shader with user-defined varying input/output).
The type layout logic added cases for handling raytracing payload and hit-attribute data, but this is currently just a stub implementation that follows the same logic as for varying `in` and `out` parameters (it cannot give meaningful byte sizes/offsets right now).
To my knowledge the GLSL spec doesn't currently specify anything about layout, and I haven't read the DXR spec language carefully enough to know what it says about layout.
A future change should update the layout logic to allow for byte-based layout of ray payloads, etc. so that we can query this information via reflection.
The GLSL legalization logic in `ir.cpp` was updated to factor out the per-entry-point-parameter code into its own function, and then that function was updated to special-case the input/output of a ray-tracing shader.
While for rasterization stages we typically want to take the user-declared input/output and "scalarize" it for use in GLSL (in part to deal with language limitations, and in part to tease system values apart from user-defined input/output), the GLSL spec for raytracing requires payload and hit attribute parameters to be declared as single variables. There is also the issue that even for an `in out` parameter, a ray payload parameter should only turn into a single global, whereas the handling for varying `in out` parameters generates both an `in` and an `out` global for the GLSL case.
Other than the handling of entry point parameters, the GLSL legalization pass doesn't need to do anything special for ray tracing shaders.
The trickiest change in the `emit.cpp` logic is that we now generate `location`s for ray payload arguments (the outgoing from a `TraceRay()` call) on demand during code generation.
This is a bit hacky, and it would be nice to handle it as a separate pass on the IR rather than clutter up the emit logic, but this approach was expedient.
Basically, any of the global variables that got generated from the `static` declarations in the standard library implementation of `TraceRay()` will trigger the logic to assign them a `location`.
The logic for emitting intrinsic operations added a few new `$`-based escape sequences. The `$XP` case handles emitting the location of a generated ray payload variable; this is how we emit the matching location at the site where we call `traceNVX`. The `$XT` case emits the appropriate translation for `RayTCurrent()` in HLSL, because it maps to something different depending on the target stage.
All of the test cases here consist of a pair of an HLSL/Slang shader written to the DXR spec, plus a matching GLSL shader for a baseline.
The GLSL shaders are carefully designed so that when fed into glslang they will produce the same SPIR-V as our cross-compilation process.
This kind of testing is quite fragile, but it seems to be the best we can do until our testing framework code supports *both* DXR and VKRay.
A bunch of the core changes ended up being blocked on issues in the rest of the compiler, so some additional features go implemented or fixed along the way:
The first big wall this work ran into was that the `__specialized_for_target` modifier hasn't actually been working correctly for a while.
It turns out that for the one function that is using it, `saturate()`, we have been outputting the workaround GLSL function in *all* cases (including for HLSL output) rather than only on GLSL targets.
The problem here is that for a generic function with a `__specialized_for_target` modifier or a `__target_intrinsic` modifier, the IR-level decoration will end up attached to the `IRFunc` instruction nested in the `IRGeneric`, but the logic for comparing IR declarations to see which is more specialized (via `getTargetSpecializationLevel()`) was looking only at decorations on the top-level value (the generic).
The quick (hacky) fix here is to make `getTargetSpecializationLevel()` try to look at the return value of a generic rather than the generic itself, so that it can see the decorations that indicate target-specific functions.
A more refined fix would be to attach target-specificity decorations to the outer-most generic (to simplify the "linking" logic).
The only reason not to fold that into the current fix is that the `__target_intrinsic` modifier currently serves double-duty as a marker of target specialization *and* information to drive emit logic. The latter (the emit-related stuff) currently needs to live on the `IRFunc`, and moving it to the generic could easily break a lot of code.
This needs more work in a follow-on fix, but for now target specialization should again be working.
The other big gotcha that the simple "just use the standard library" strategy ran into was that function-`static` variables weren't actually implemented yet, and in particular function-`static` variables inside of generic functions required some careful coding.
The logic in `lower-to-ir.cpp` has this `emitOuterGenerics()` function that is supposed to take a declaration that might be nested inside of zero or more levels of AST generics, and emit corresponding IR generics for all those levels.
This is needed because two different AST functions nested inside a single generic `struct` declaration should turn into distinct `IRFunc`s nested in distinct `IRGeneric`s.
The tricky bit to making that all work is that the same AST-level generic type parameter will then map to *different* IR-level instructions (the parameters of distinct `IRGeneric`s) when lowering each function.
The existing logic handled this in an idiomatic way by making "sub-builders" and "sub-contexts."
This change refactors some of the repeated logic into a `NestedContext` type to help simplify the pattern, and applies it consistently throughout the `lower-to-ir.cpp` file.
Besides that cleanup, the major change is `lowerFunctionStaticVarDecl` which, unsurprisingly, handles lower of function-`static` variables to IR globals.
The careful handling of nested contexts here is needed because if we are in the middle of lowering a generic function, then a `static` variable should turn into its *own* `IRGeneric` wrapping an `IRGlobalVar`. The body of the function should refer to the global variable by specializing the global variable's `IRGeneric` to the parameters of the *functions* `IRGeneric`. This tricky detail is handled by `defaultSpecializeOuterGenerics`.
An additional subtlety not actually required for this raytracing work (and thus not properly tested right now) is handling function-`static` variables with initializers.
These can't just be lowered to globals with initializers, because HLSL follows the C rule that function-`static` variables are initialized when the declaration statement is first executed (and this could be visible in the presence of side-effects).
The lowering strategy here translates any `static` variable with an initializer into *two* globals: one for the actual storage, plus a second `bool` variable to track whether it has been initialized yet.
There are some opportunities to optimize this case, especially for `static const` data, but that will need to wait for future changes.
We've slowly been shifting away from the model where a user thinks of a "profile" as including both a stage and a feature level.
Instead, the user should think about selecting a profile that only describes a feature level (e.g., `sm_6_1`, `glsl_450`, etc.), and then separately specifying a stage (`vertex`, `raygeneration, etc.) for each entry point.
The challenge here is that the command-line processing still only had a single `-profile` switch, and no way to specify the stage.
Adding the `-stage` option was relatively easy, but making it work with the existing validation logic for command-line arguments was tricky, because of the complex model that `slangc` supports for compiling multiple entry points in a single pass.
* In `slang.h` add new reflection parameter categories for ray payloads and hit attributes, as part of entry point input/output signatures.
* A previous change already updated our copy of glslang to one that supports the `GL_NVX_raytracing` extension, so in `slang-glslang.cpp` we just needed to map Slang's `enum` values for the raytracing stage names to their equivalents in the glslang code.
* Moved the logic for looking up a stage by name (`findStageByName()`) out of `check.cpp` and into `compiler.cpp`, with a declaration in `profile.h`
* Added a `$z` suffix to the GLSL translation of `Texture*.SampleLevel()`, to handle cases where the texture element type is not a 4-component vector. Note that this fix should actually be applied to *all* these texture-sampling operations, but I didn't want to add a bunch of changes that are (clearly) not being tested right now.
* The layout logic for entry points was updated to correctly skip producing a `TypeLayout` for an entry point result of type `void`, which meant that the related emit logic now needs to guard against a null value for the result layout.
* In `ir.cpp`, dump decorations on every instruction instead of just selected ones, so that our IR dump output is more complete.
* Added a command-line `-line-directive-mode` option so that we can easily turn off `#line` directives in the output when debugging. Not all cases where plumbed through because the `none` case is realistically the most important.
* Parser was fixed to properly initialize parent links for "scope" declarations used for statements, so that we can walk backwards from a function-scope variable (including a `static`) and see the outer function/generics/etc.
* Added GLSL 460 profile, since it is required for ray tracing. Also updated the logic for computing the "effective" profile to use to recognize that GLSL raytracing stages require GLSL 460.
* Added some conventional ray-tracing shader suffixes to the handling in `slang-test`. This code isn't actually used, but was relevant when I started by copy-pasting some existing VKRay shaders as the starting point for my testing.
* Fixup: typos
|
|
* * Change the layout of IROp such that 'main' IROps are 0-x.
* Removed MANUAL_RANGE instuction types, as no longer needed.
* Work in prog on optimizing.
* * Constant time lookup for IROpInfo
* Refactor and document a little more the IROp layout
* Mark ops that use 'other' bits
* Fix typo in definition of kIROpFlag_UseOther
|
|
* Fixes around atomic operations
Work on #651
The existing handling of atomic operations had a few issues:
* The HLSL atomic functions (`Interlocked*`) didn't have mappings to GLSL
* Atomic operations on images weren't supported at all because the subscript operation on `RWTexture*` types didn't provide a `ref` acessor
* The HLSL atomic functions were only providing the overloads that return the previous value through an `out` parameter, and not the ones that ignore the previous value.
This change fixes these issues with the following changes:
* `RWTexture*` types now have a `ref` accessor on their subscript operation which maps to a new `imageSubscript` operation in the IR. By default this translates back to `tex[idx]` in output HLSL, but it makes a custom mapping possible for GLSL
* The `Interlocked*` function definitions were expanded to include the overloads without the `out` parameter
* GLSL translations were added for the `Interlocked*` functions. These mappings use some new customization points in the intrinsic operation emit logic to support outputting calls to either `atomic*` or `imageAtomic*` as required, and to expand an argument that is a subscript into an image as multiple arguments.
This whole approach is quite hacky, and it doesn't seem like the approach we should take in the long run.
* Fix: typo in InterlockedAnd lowering
One of the cases of `InterlockedAnd` was lowering to `atomicAnd` with a `$0` where we wanted the `$A` substitution to handle the possibility of an image.
|
|
* Change the layout of IROp such that 'main' IROps are 0-x. (#649)
* Removed MANUAL_RANGE instuction types, as no longer needed.
|
|
* * Added support for strings in IR with IRStringLit - with storage of chars after it
* Added kIRDecorationOp_Transitory - can be used for detecting instructions constructed on stack
* Made IRConstant hashing work off type
* Fix comment that is out of date about how an instruction is determines to hold a transitory string.
|
|
Fixes #627
The front-end has support for `RasterizerOrderedBuffer` and `RasterizerOrderedTexture*`, but left out support for:
* `RasterizerOrderedByteAddressBuffer`
* `RasterizerOrderedStructuredBuffer`
[Nitpick: these tyeps are all amazingly annoying to type. It is easy to want to write `RasterOrdered` instead of the bulkier `RasterizerOrdered`, and almost everybody does in casual speech. There's already the issue of wanting to type `StructureBuffer` (a buffer of structures) instead of `StructuredBuffer` (a buffer that is... structured?). Then you have `ByteAddressBuffer` which is just adding to the confusion because it is nominally a "byte addressable" buffer (so that `ByteAddressedBuffer` would actually make sense), but then actually *isn't* byte addressable in practice.]
There were a few `TODO` comments related to this already, and this change was mostly a matter of doing a find-in-files for `RWByteAddressBuffer` and `RWStructuredBuffer` and adding matching `RasterizerOrdered` cases.
The test I added just checks that these types make it through the front-end, and doesn't do any actual confirmation that they work as intended. It is worth noting that the handling of ordering in GLSL/VK is different from in HLSL ("pixel shader interlock" instead of "rasterizer ordered views"), so coming up with a cross-compilation story would need to be a later step.
|
|
* Fix typo OuptutTopologyAttribute -> OutputTopologyAttribute
First pass support for handing tesselation shaders - domain and hull.
* Added attribute PatchConstantFuncAttribute
* Added visitHLSLPatchType(HLSLPatchType* type) such that the patch type template parameters are handled
* Added IRNotePatchConstantFunc - such that the patch constant function is referenced within IR
* Added support for outputing typical tesselation attributes (although minimal validation is performed)
* Added findFunctionDeclByName
* Small improvements to diagnostic.
* Improved diagnostics and checking for geometry shader attributes.
* Added diagnostic if patchconstantfunc is not found
Handle assert failure when outputing a domain shader alone and therefore attr->patchConstantFuncDecl is not set.
* Simple script tess.hlsl to test out domain/hull shaders.
* Added url for where hull shader attributes are defined.
* Fix unsigned/signed comparison warning.
* Restore removal of fix in "Improve generic argument inference for builtins (#598)"
* Update tessellation test case to compare against fxc
The test was previously comparing against fixed expected DXBC output, but this caused problems when the test runner tried to execute the test on Linux (where there is no fxc to invoke...), and would also be a potential source of problems down the road if different users run using different builds of fxc.
The simple solution here is to convert the test to compare against fxc output generated on the fly. That test type is already filtered out on non-Windows builds, so it eliminates the portability issue (in a crude way).
I also changed the test to compile both entry points in one compiler invocation, just to streamline things into fewer distinct tests.
* Eliminate unnecessary call to `lowerFuncDecl`
In a very obscure case this could cause a bug, if the patch-constant function had somehow already been lowered (because it was called somewhere else in the code).
The call should not be needed because `ensureDecl` will lower a declaration on-demand if required, so eliminating it causes no problems for code that wouldn't be in that extreme corner case.
|
|
Fixes #581
This change adds a new parameter passing mode `__ref` to exist alongisde `in`, `out`, and `inout`.
The `__ref` modifier indicates true by-reference parameter passing (whereas `inout` is copy-in-copy-out).
This is not intended to be something that users interact with directly, but rather a low-level feature that lets us provide a correct signature for the `Interlocked*()` operations in the standard library.
Most of the support for passing what are logically addresses around already exists in the IR, so the majority of the work here is just in introducing the new type `Ref<T>` and then using it appropriately when lowering `__ref` parameters/arguments to the IR.
|
|
This was a known issue in our IR representation, which was now biting a user. The basic problem is that in code like the following:
```hlsl
RWStructureBuffer<float4> buffer;
...
buffer[index].xz = value;
```
we ideally want to be able to reproduce the original HLSL code exactly, but that requires directly encoding the way that this code writes to two elements of a vector, but not the others.
The currently lowering strategy we had produced IR something like:
```hlsl
float4 tmp = buffer[index];
tmp.xz = value;
buffer[index] = tmp;
```
That transformation might seem valid, but it has some big problems:
* It generates UAV reads that are not needed, which could impact performance
* It performs read-modify-write operations on memory that the programmer didn't explicitly write, which could create data races
The fix here is somewhat obvious: if the "base" of a swizzle operation on a left-hand side resolves to a pointer in our IR, then we can output a "swizzled store" instead of the read-modify-write dance. We currently keep the read-modify-write around since it is potentially needed as a fallback in the general case.
Along the way I also tried to make sure that we handle the case where we have a swizzle of a swizzle on the left-hand side:
```hlsl
buffer[index].xz.y = value;
```
That code should behave the same as `buffer[index].z = value`. I am currently detecting and cleaning up this logic in the lowering path for `SwizzleExpr`, because that is the only place in the lowering logic that "swizzled l-values" currently get created.
|
|
* Introduce an IR-level type system
Up to this point, the Slang IR has used the front-end type system to represent types in the IR.
As a result (but ultimately more importantly) the IR representation of generics and specialization has used AST-level concepts embedded in the IR.
For example, to express the specialization of `vector<T,N>` to a concrete type `float` for `T`, we needed an IR operation that could represent the specialization, with operands that somehow represented the type argument `float`.
The whole thing was very complicated.
The big idea of this change is to introduce a new representation in which types in the IR are just ordinary instructions, so that using them as operands makes sense. The hierarchy of IR types closely mirrors the AST-side hierarchy for now, and that will probably be something we should maintain going forward.
In order to make these changes work, though, I also had to do major overhauls of things like the way substitutions are performed, how we check interface conformances, the way lookup through interface types is done, etc. etc. This is a big change, and unfortunately any attempt to summarize it in the commit message wouldn't do it justice.
* Fix 64-bit build warning
* Fix up some clang warnings/errors
|
|
The main practical change here is that things that used to be `IRValue`s, like literals, are now being expressed as instructions in the global scope.
In order to validate that things are actually being handled correctly, this change introduces an explicit "validation" pass that can be run on the IR to check for different invariants (although it doesn't check many of the important ones right now). I've left the validation pass turned off by default, but with a command-line flag to enable it. We may want to make it be on by default in debug builds, just to keep us honest. The main invariant for the moment is that when on IR instruction is used as an operand to another, it had better come from the same IR module.
Some of the existing passes were violating this rule, in particular when it came to cloning of witness tables related to global generic parameter substitution. Those features can in theory be handled better now by allowing `specialize` instructions at other scopes, but I didn't want to over-complicate this change, so I make just enough fixes to ensure that these steps always clone witness tables they get from the "symbols" on an IR specialization context. In order for this to work when recursively specializing, I had to ensure that the logic for generic specialization had a notion of a "parent" specialization context that it would fall back to to perform cloning when necessary.
This change keeps the logic that was caching and re-using the instructions for literal values within a module, but adds some logic that isn't really being tested right now for picking the right parent instruction to insert a constant instruction into. This logic doesn't trigger right now because all of the cases we are using it on have zero operands (and so they always get "hoisted" to the global scope), but eventually for things like types we want to be able to support instructions with operands (e.g., `vector<float, 4>`) and handle the case where some of those operands come from different scopes (e.g., when nested inside a generic).
The final change here is mostly cosmetic: the `IRBuilder` is now more abstract about where insertion occurs: it tracks a single `IRParentInst` to insert into, and then an optional `IRInst` to insert before. In the common case, that parent is an `IRBlock`, but it could conceivably also be the global scope, or a witness table, etc. Use sites where we used to change those fields directly now use distinct methods `setInsertInto(parent)` and `setInsertBefore(inst)` which capture the two cases we care about. Accessors are also defined to extract the current block (if the current parent is a block), and the current "function" (global value with code, if the current parent is a global value with code, or a block inside one).
With this work in place, it should be possible for a follow-on change to start putting `specialize` instructions at the global scope and thus clean up some of the on-the-fly specialization work. This work should also help with some of the requirements around a distinct IR-level type system and more explicit generics.
|
|
* IR: "everything is an instruction"
This change tries to streamline the representation of the IR in the following ways:
* Every IR value is an instruction (there is no `IRValue` type any more)
* All IR values that can contain other values share a single base (`IRParentInstruction`)
* Dynamic casts to specific IR instruction types can be accomplished with a new `as<Type>(inst)` operation, that uses the IR opcode to implement casts.
The biggest change in terms of number of lines is getting rid of `IRValue`. The diff here could probably be smaller if I'd just done `typedef IRInst IRValue;`.
Along the way I also renamed the `getArg`/`getArgs`/`getArgCount` combination over to `getOperand`/`getOperands`/`getOperandCount` to avoid being confusing when we have something like a `call` instruction where the "arguments" of the call don't line up with the operands of the instruction.
I also tried to clean up the representation of lists of child instructions to try to make it easier to iterate over them with C++ range-based `for` loops. Developers still need to be careful about mutating the contents of a block while iterating over it in this fashion (e.g. if you remove the "current" element, the iteration will end prematurely).
Probably the thorniest change here is that parameters are now just represented as the first N instructions in a block, which means:
* We need to perform a linear search to find the end of the parameter list. This is probably not often a problem, because usually you would be iterating over the parameters anyway, and that will be linear in the number of parameters.
* Algorithms that iterate over a block either need to ignore parameters, treat parameters just like other instructions, or somehow cleave the list into the range of parameters, and the range of "ordinary" instructions (which involves the same linear search above).
* When inserting into a block, we need to be careful not to insert instructions at invalid locations (e.g., insert a temporary before the parameters, or insert a parameter in the middle of the code). I can't pretend that I've handled the details of that here. (This is no different than having to make the same adjustments for phi nodes in a typical SSA representation)
* One possible future-proof approach is to implement a pass that sorts the instructions in a block so that parameters always come first. That would let us implement passes without caring about this detail, and then clean up right before any pass that cares about the relative order of parameters and other instructions.
The current change is missing any work to make literals and other instructions that used to be `IRValue`s properly nest inside of their parent module. Right now these instructions are just left unparented, and may actually end up being shared between distinct modules. Fixing that will need a follow-up change. The biggest challenge there is that it introduces instructions at the global scope that aren't `IRGlobalValue`s.
This change doesn't try to take advantage of any of the new flexibility (e.g., by nesting `specialize` instructions inside of witness tables). The goal is to do exactly what we were doing before, just with a different representation.
* Warning fix
|
|
* Basic IR support for `static const` globals
Our strategy for lowering global *variables* can fall back to putting their initialization into a function, but that isn't really appropriate for global constants (it also isn't appropriate for arrays, but we'll need to deal with that seaprately).
This change adds a distinct case for global constants (rather than treating them as variables), and forces the emission logic to always emit them as a single expression.
Doing this makes assumptions about how the IR for these constants gets emitted (and what optimziations might do to it).
In order to make things work, I had to switch the handling of initializer-list expressions to not be lowered via temporaries and mutation (since that isn't a good fit for reverting to a single expression).
I've added a single test case to ensure that this works in the simplest scenario. My next priority will be to see if this unblocks my work in Falcor.
* Fixup: bug fixes
|
|
* Re-define deprecated compile flags
By including these flags in the header file, with a value of zero, we can allow some existing code to compile even after the major changes to the implementation.
* The `SLANG_COMPILE_FLAG_NO_CHECKING` option will effectively be ignored, since checking is always enabled.
* The `SLANG_COMPILE_FLAG_SPLIT_MIXED_TYPES` option will now act as if it is always enabled (and indeed some of the code has been relying on this flag being set always).
* Make subscript operators writable for writable textures
This even had a `TODO` comment saying that we needed to fix it, and now I'm seeing semantic checking failures because we didn't define these and so we find assignment to non l-values.
* Fix definitions of any() and all() intrinsics
These should always return a scalar `bool` value, but they were being defined wrong in two ways:
1. They were using their generic type parameter `T` in the return type
2. They were returning a vector in the vector case, and a matrix in the matrix case.
This change just alters the return type to be `bool` in all cases.
* Fix bug in SSA construction
When eliminating a trivial phi node, it is possible that the phi is still recorded as the "latest" value for a local variable in its block.
When later code queries that value from the block (which can happen whenever another block looks up a variable in its predecessors), it would get the old phi and not the replacement value.
I simply added a loop that checks if the value we look up is a phi that got replaced, and then continues with the replacement value (which might itself be a phi...). A more advanced solution might try to get clever and have the map itself hold `IRUse` values so that we can replace them seamlessly.
* Simplify IR control flow representation
This change gets rid of various special-case operations for conditional and unconditional branches, and instead requires emit logic to recognize when a direct branch is targetting a `break` or `continue` label.
The new approach here isn't perfect, but it seems beter than what we had before, because it can actually work in the presence of control-flow optimizations (including our current critical-edge-splitting step).
* Load from groupshared isn't groupshared
When loading from a `groupshared` variable, the resulting temporary shouldn't have the `groupshared` qualifier on it.
This might eventually need to generalize to a better understanding of storage modifiers in the IR, but I don't really want to deal with that right now.
* Don't emit references to typedefs in output code
Now that we are using the IR for all codegen, we shouldn't be dealing with surface-level things like `typedef` declarations in the output code; just use the type that was being referred to in the first place.
* Fix floating-point literal printing for IR
The IR was calling `emit()` instead of `Emit()` (we really need to normalize our convention here), and was implicitly invoking a default constructor on `String` that takes a `double` (that constructor should really be marked `explicit`), and which doesn't meet our requirements for printing floating-point values.
* Fix error when importing module that doesn't parse
We already added a case to bail out if semantic checking fails, but neglected to add a case if there is an error during parsing of a module to be imported.
Note: this logic doesn't correctly register the module as being loaded (but still in error), so users could see multiple error messages if there are multiple `import`s for the same module.
* Improve error message for overload resolution failure
- Drop debugging info from the candidate printing
- Add cases to print `double` and `half` types properly
* Fixup: switch loopTest to ifElse in expected IR output
|
|
* Generate SSA form for IR functions
The basic idea here is simple: in the front-end after we have lowered the AST to initial IR we will apply a set of "mandatory" optimization passes. The first of these is to attempt to translate the all functions into SSA form so that they are amenable to subsequent dataflow optimizations. Eventually, the mandatory optimization passes would include diagnostic passes that make sure variables aren't used when undefined, etc.
Just doing basic SSA generation already cleans up a lot of the messiness in our IR today, because constructs that used to involve many local variables can now be handled via SSA temporaries.
The implementation of SSA generation is in `ir-ssa.cpp`, and it follows the approach of Braun et al.'s "Simple and Efficient Construction of Static Single Assignment Form." I used this instead of the more well-known Cytron et al. algorithm because Braun's algorith mis very simple to code, and does not require auxiliary analyses to generate the dominance frontier.
The main wrinkle in our SSA representation right now is that instead of using ordinary phi nodes, we instead allow basic blocks to have parameters, where predecessor blocks pass in different parameter values. This encodes information equivalent to traditional phi nodes, but has two (small) benefits:
1. There is no fixed relationship between the order of phi operands and predecessor blocks, so we don't have to worry about breaking the phis when we alter the order in which predecessors are stored. This is important for us because predecessors are being stored implicitly.
2. It is easy to operationalize a "branch with arguments" either when lowering to other languages, or when interpreting the IR. A branch with arguments is implemented as a sequence of stores from the arguments to the parameters of the target block (very similar to a call), followed by a jump to the block.
Relevant to the above, this change also adds an interface for enumerating the predecessors or successors of a block in our CFG. Rather than use an auxliary structure, we directly use the information already encoded in the IR:
* The sucessors of a block are the target label operands of its terminator instruction. In our IR this is a contiguous range of `IRUse`s, possible with a stride (to account for the way `switch` interleaves values and blocks).
* The predecessors of a block are a subset of the uses of the block's value. Specifically, they are any uses that are on a terminator instruction, and within the range of values that represent the successor list of that instruction.
One important limitation of the "blocks with arguments" model for handling phis is that it is really only convenient to stash extra arguments on an unconditional terminator instruction. This change works around this prob lem by breaking any "critical edges" - edges between a block with multiple successors and one with multiple predecessors. We assume that "phi" nodes will only ever be needed on a block with multiple predecessors, and because critical edges are broken, each of these predecessors will then have only a single successor, so its branch instruction can handle the extra arguments.
This change introduces a notion of an "undefined" instruction in the IR. This is handled as an instruction rather than a value because I anticipate that we will want to distinguish different undefined values when it comes time to start issuing error messages (those messages will need to point to the variable that was used when undefined).
* Fix expected test output.
Another change was merged that enabled the `glsl-parameter-blocks` test, and its output is affected by our IR optimization work.
|
|
- support overloaded generic function. this involves adding a new expression type, `OverloadedExpr2` to hold the candidate expressions for the generic function decl being referenced.
- make BitNot a normal IROp instead of an IRPseudoOp
- make sure we clone the decorations of parameters when cloning ir functions
- propagate geometry shader entry point attributes (`[maxvertexcount]` and `[instance]`) through HLSL emit
- IR emit: handle geometry shader entry-point parameter decorations, such as 'triangle'.
- IR emit: treat geometry shader stream output typed ir value as `should fold into use`.
|
|
|
|
* IR: fixes for subscript accessors
Fixes #320
This is a bunch of fixes for handling of `__subscript` operations on builtin types (notably `RWStructuredBuffer` and `StructuredBuffer` at this point).
- Automatically add a `GetterDecl` to any subscript decalratio was declithout any accessors. This avoids hitting a null- dereference in the emit logic.
- Add a notion of a `RefAccessor` (declared with `ref`) as a peer to getters and setters. The idea is that a `ref` accessor returns a pointer to the element data, so that it can be used for both getting and setting values. This is closer to the behavior of `RWStructuredBuffer` element access in HLSL.
- Fixes for dealing with "access chains" where there might be a combination of a subscript (where the is a `get` and `set` but no `ref`) and member access, so that we have to read the base value into a temp, modify it, and then write it back.
- This logic is still a bit of a mess, so we will eventually want to take a more consistent pass over this to deal with how we "materialize" values for setters.
- Update `RWStructuredBuffer` to have a `ref` accessor, and then fix up the IR tests to handle the new opcode that I added for it.
- Note: I didn't handle this as an intrinsic simply because the `tests/ir/*` tests aren't really set up to handle builtins with ugly mangled names.
* Fixup: type error in VM for buffer element ref
I was using the result type of the op as the element type for computing the element address, but the result type is a pointer to the real element type.
This caused test failures on 64-bit platforms, where the stride of the buffer in the `ir/factorial` test needs to be 4.
The fix is to assume the result type is a pointer, and extract the pointed-to type out of that.
|
|
- Change function mangling so we use `p<parameterCount>p` instead of just `p<parameterCount>` to avoid the parameter count running into digits at the start of a mangled type name and tripping up the un-mangling logic.
- We really need to step back at some point and define our mangling scheme a bit more carefully, especially if we are going to keep going down this road where un-mangling things is important for generating HLSL output.
- Also allow the unmangling logic to unmangle a few more cases of generic parameters, so that it can skip over them to get to the parameter count of the underlying function.
- Add a notion of an `unreachable` instruction to the IR, and emit it as the terminator (if needed) at the end of the last block for a function with a non-void return type.
- This does *not* implement any logic to emit a diagnostic if the `unreachable` turns out to be potentially reachable
- Fix a bug in IR specialization of generics where we can't create two different specializations of the same function, because both get registered in the same hash map
With all these fixes, testing in Falcor modified to use the full Slang compiler and IR for all HLSL/Slang:
- The UI and text rendering shaders yield HLSL that compiles without error; no idea if they actually *work*
- The ModelViewer shaders yield HLSL, but there are some issues (looks like type legalization isn't applying to stuff inside constant buffers)
|
|
* IR: add support for `switch` statements
Fixes #273
This is just something we hadn't gotten to yet on the IR. The actual design of the instruction is unsurprising (once you take into consideration the requirement for structured control flow). A `switch` instruction takes the form:
switch <condition> <breakLabel> <defaultLabel> [<caseVal> <caseLabel>]*
Where `condition` is the value to switch on, `breakLabel` is the "join point" after the original `switch` statement, `defaultLabel` is where to go if the value doesn't match any case, and each pair of `caseVal` and `caseLabel` is what to do on a particular value. It is required that `caseVal` be a literal, but this isn't currently being enforced in the IR (the front-end should be making a check and constant-folding the case labels).
For structured control flow, we also make the assumption that the cases are in order: cases with the same label must be grouped together, and any case that falls through to another must come right before it.
Given this representation, the emit logic can reconstruct a `switch` statement with relative ease, given the machinery we already have. It makes sure to group together case values with the same label (again, assuming they are contiguous), and will insert the `default:` label in with whatever group it belongs to.
Actually emitting code for a `switch` statement seems superficially simple, until you realize that a complete implementation needs to handle stuff like "Duff's Device." The current implementation makes the assumption that all `case` and `default` statements are directly nested under a `switch`, and that there is no way for control flow to enter a case except by the `switch` itself, or fall-through.
In order to facilitate the grouping of cases in the IR-to-HLSL emit logic, the AST-to-IR lowering logic tries to detect cases where there are multiple `case`s in a row, and emit only a single label for them.
One big/annoying gotcha is that we don't properly handle the case where a `default:` case has a non-trivial fall-throguh to another case. That seems fine for now since HLSL doesn't support fall-through anyway, but it probably needs to get detected somewhere in the Slang compiler (e.g., maybe we should add a diagnostic pass over the IR that detects target-specific problems like that and emits errors).
* IR: Add support for empty statements.
- Add empty statement in `lower-to-ir.cpp`
- Go ahead and eliminate the statement catch-all and explicitly enumerate the cases we don't support
- Fix up parser for block statements so that it doesn't leave a null statement as the body of a `{}`
- Add an empty statement to one of the cases for the `switch` test, to ensure we are testing empty statements
|
|
- Add definition of `discard` instruction
- A `discard` is a terminator instruction, just like `returnVoid`
- Lower `DiscardStmt` in AST to a `discard` instruction in the IR
- Emit `discard` instruction as a `discard;` statement when emitting HLSL/GLSL
- Add a test case using the "graphics compute" mode that tests discard. The test writes to one entry in a UAV before doing a conditional (always true at runtime) discard, and then writes to another entry; we expect to see the results of the first write, but not the second.
|
|
This change includes a lot of infrastructure work, but the main point is to allow code like the following:
```
// define an interface
interface Helper { float help(); }
// define a generic function that uses the interface
float test<T : Helper>( T t ) { return t.help(); }
// define a type that implements the interface
struct A : Helper { float help() { return 1.0 } }
// define an ordinary function that calls the
// generic function with a concrete type:
float doIt()
{
A a;
return test<A>(a);
}
```
Getting this to generate valid code involves a lot of steps. This change includes the initial version of all of these steps, but leaves a lot of gaps where more complete implementation is required.
The changes include:
- Member lookup on types has been centralized, and now handles the case where the type we are looking for a member in is a generic parameter (e.g., given `t.help()` we can now look up `help` in `Helper` by knowing that `t` is a `T` and `T` conforms to `Helper`).
- There is an obvious cleanup still to be done here where the same exact logic should be used to look up available "constructor" declarations inside a type when the type is used like a function.
- Add a notion of subtype constraint "wittnesses" to the type system. When a generic is declared as taking `<T : Helper>` it really takes two generic parameters: the type `T` and a proof that `T` conforms to `Helper`. The actual arguments to a generic will then include both the type argument and a suitable witness argument (both type-level values).
- As it stands right now, a witness wraps a `DeclRef` to the declaration that represents the appropriate subtype relationship. So if we have `struct A : Helper`, that `: Helper` part turns into an `InheritanceDecl` member, and a reference to that member can serve as a witness to the fact that `A` conforms to `Helper`.
- Make explicit generic application `G<A,B>` synthesize the additional arguments that represent conformances required by the generic.
- This does *not* yet deal with the case where a generic is implicitly specialized as part of an ordinary call `G(a,b)`
- A bug fix to not auto-specialize generics during lookup. The problem here was related to an attempted fix of an earlier issue.
During checking of a method nested in a generic type, we were running into problems where `DeclRefType::create()` was getting called on an un-specialized reference to `vector`, and this was leading to a crash when the code looked for the arguments for the generic. This was worked around by having name lookup automatically specialize any generics it runs into while going through lookup contexts.
That choice creates the problem that in a generic method like this:
```
void test<T>(T val) { ... }
```
any reference to `val` inside the body of `test` will end up getting specialized so that it is effectively `test<T>::val`, when that isn't really needed.
- Add front-end logic to check that when a type claims to conform to an interface it actually must provide the methods required by the interface. The checking process goes ahead and builds a front-end "witness table" that maps declarations in the interface being conformed to over to their concrete implementations for the type.
- At the moment the checking is completely broken and bad: it assumes that *any* member with the right name is an appropriate declaration to satisfy a requirement. That obviously needs to be fixed.
- Add an explicit operation to the IR for lookup of methods: `lookup_interface_method(w, r)` where `w` is a reference to the "witness" value and `r` is an `IRDeclRef` for the member we want to look up.
- Add an explicit notion of witness tables to the IR. These end up being the IR representation of an `InheritanceDecl` in a type, and they are generated by enumerating the members that satisfy the interface requirements (which were handily already enumerated by the front-end checking). The witness table is an explicit IR value, and so it will be referenced/used at the site where conformance is being exploited (e.g., as part of a `specialize` call), so it should be safe to eliminate witness tables that are unused (since they represent conformances that aren't actually exploited). Similarly, the entries in a witness table are uses of the functions that implement interface methods, and so keep those live.
- In order to implement the above, I did a bit of a cleanup pass on the IR representation so that there is an `IRUser` base that `IRInst` inherits from, so that we can have users of values that aren't instructions.
- One annoying thing is that because of how types and generics are handled in the IR, we needed a way to have a type-level `Val` that wraps an IR-level value: e.g., to allow an IR-level witness table to be used as one of the arguments for specialization of a generic. The design I chose here is to have a "proxy" `Val` subclass (`IRProxyVal`) that wraps an `IRValue*`. These should only ever appear as part of types and `DeclRef`s that are used by the IR.
- One annoying bit here is that an IR value might then have a use that is not manifest in the set of IR instructions, and instead only appears as part of a type somewhere.
- I'm not 100% happy with this design, but it seems like we'd have to tackle similar issues if/when we eventually allow functions to have `constexpr` or `@Constant` parameters
- Make generic specialization also propagate witness table arguments through to their use sites (this is mostly just the existing substitution machinery, once we have `IRProxyVal`), and then include logic to specialize `lookup_interface_method` instructions when their first operand is a concrete witness table.
All of this work allows a single limited test using generics with constraints to pass, but more work is needed to make the solution robust.
|
|
There are two big changes here:
- Add logic during the initial IR cloning pass for an entry point + target that tries to pick the best possible version of any target-overloaded function. This allows us to pick the intrinsic version of `saturate()` when compiling for HLSL output, but then pick the non-intrinsic version (that is implemented in terms of `clamp()`) when targetting GLSL.
- Add an initial specialization pass that tries to deal with generics. This required some fixing work to IR generation, so that we correctly generate explicit operations to specialize a generic for specific types (this is currently implemented as a `specialize` instruction that takes the generic to specialize plus a declaration-reference that represents the specialized form). With that work in place, we can scan for `specialize` instructions inside of non-generic functions, and use them to trigger generation of specialized code. We rely on the name-mangling scheme to help us find pre-existing specializations when possible.
There are also a bunch of cleanups encountered along the way:
- Don't use the explicit `layout(offset=...)` for uniforms, because it isn't supported by all current drivers. For now we will just assume that our layout rules compute the same values that the driver would for un-marked-up code. We can come back later and try to implement a workaround in the cases where this doesn't apply (e.g., by re-running the layout logic as part of emission, and dropping layout modifiers from variables that don't need explicit layout).
- Fix some issues in IR dump printing so that we print function declarations more nicely.
- Testing: print out failing pixel when image-diff fails
|
|
* IR: overhaul IR design/implementation
Closes #192
Closes #188
This is a major overhaul of how the IR is implemented, with the primary goal of just using the AST-level type representation as the IR's type representation, rather than inventing an entire shadow set of types (as captured in issue #192).
One consequence of this choice is that types in the IR are no longer explicit "instructions" and are not represented as ordinary operands (so a bunch of `+ 1` cases end up going away when enumerating ordinary operands).
Along the way I also got rid of the embedded IDs in the IR (issue #188) because this wasn't too hard to deal with at the same time.
Another related change was to split the `IRValue` and `IRInst` cases, so that there are values that are not also instructions. Non-instruction values are now used to represent literals, references to declarations, and would eventually be used for an `undef` value if we need one. IR functions, global variables, and basic blocks are all values (because they can appear as operands), but not instructions.
The main benefit of this approach is that the top-level structure of a bytecode (BC) module is much simpler to understand and walk, and BC-level types are represented much more directly (such that we could conceivably use them for reflection soon).
* fixup: 64-bit build fix
* fixup: try to silence clang's pedantic dependent-type errors
* fixup: bug in VM loading of constants
|
|
None of these changes are made "live" at the moment. I'm just trying to get them checked in to avoid divering too far from `master` at any point during development.
- Add basic emit logic to produce GLSL from the IR in a few cases (the existing IR emit logic was ad hoc and HLSL-specific)
- When lowering a function declaration, walk up its chain of parent declarations to collect additional parameters as needed
- When lowering a call, make sure to add generic arguments that come from the declaration reference being called
- Attach a "mangled name" to symbols when lowering, so that we can eventually use that name to resolve things for linkage.
- After the above work, I had to apply some fixups to make sure that generic arguments *don't* get added when the user is calling an `__intrinsic_op` function, since those should map 1-to-1 down to instructions with just their ordinary parameter list.
A big open question right now is whether I should continue to represent the generic arguments as just part of the ordinary argument list for a function, or split them out into separate `applyGeneric` and `apply` steps.
A strongly related question is whether a declaration with generic parameters should lower into a single declaration, or one declaration nested inside an outer generic declaration.
A good future step at this point would be to eliminate a lot of the `__intrinsic_op` stuff in favor of having the builtin functions include their own definitions, which might be in terms of a new expression-level construct for writing inline IR operations. This can't be done until the existing AST-to-AST path is no longer needed for cross-compilation purposes.
More immediate next steps here:
- We need a way to round-trip calls to external declaration that get handled by this mangled-name logic. Basically, if we are asked to output HLSL and we see a call to `_S...GetDimensions...(float4, t, a, ...)` we need to be able to walk the mangled name and get back to `t.getDimensions(a, ...)` without a whole lot of manual definitions to make things round-trip.
- In the other case, where a declaration isn't built-in for the chosen target, we need to be able to load a module of target-specific definitions (which will somehow map back to symbols with certain mangled names) and then look these up (by mangled name) and then load/link/inline them into the user's IR to satisfy requirements in their code.
|
|
* IR: handle control flow constructs
This change includes a bunch of fixes and additions to the IR path:
- `slang-ir-assembly` is now a valid output target (so we can use it for testing)
- This uses what used to be the IR "dumping" logic, revamped to support much prettier output.
- A future change will need to add back support for less prettified output to use when actually debugging
- IR generation for `for` loops and `if` statements is supported
- HLSL output from the above control flow constructs is implemented
- Revamped the handling of l-values, and in particular work on compound ops like `+=`
- Add basic IR support for `groupshared` variables
- Add basic IR support for storing compute thread-group size
- Output semantics on entry point parameters
- This uses the AST structures to find semantics, so its still needs work
- Pass through loop unroll flags
- This is required to match `fxc` output, at least until we implement
unrolling ourselves.
* Fixup: 64-bit build issues.
* fixup for merge
|
|
- Add support for matrix types in IR/codegen
- Add support for basic indexing operations in IR/codegen
|
|
The main interesting change here is around support for lowering of calls to "subscript" operations (what a C++ programmer would think of as `operator[]`).
An important infrastructure change here was to add an explicit AST-node representation for a "static member expression" which we use whenever a member is looked up in a type as opposed to a value. The implementation of this probably isn't robust yet, but it turns out to be important to be able to tell such cases apart.
|
|
The code previously had an enumerated type for "intrinsic" operations, and allowed functions to be marked `__intrinsic_op(...)` to indicate the operation they map to.
The nature of the IR meant that each of these intrinsic ops had to have a corresponding IR opcode, but the `enum` types weren't the same.
This change cleans things up a bit by deciding that the `__intrinsic_op(...)` modifier names an actual IR opcode, and so the `IntrinsicOp` enum is gone.
The biggest source of complexity here is that there are certain operations that need to be "intrinsic"-ish for the purposes of the current AST-based translation path, because we need them to round-trip from source to AST and back.
Right now this is being handled by defining a bunch of "pseudo-ops" which can be used in the `__intrinsic_op` modifier, but which are *not* meant to be represented in the IR.
Currently I don't actually handle this during IR generation.
In the long run, once we are using IR for everything that needs cross-compilation, we should be able to eliminate the pseudo-ops in favor of just having these be ordinary (inline) functions defined in the stdlib (e.g., the `+=` operator can just have a direct definition).
There was a second category of modifier that gets a little caught up in this, which is the `__intrinsic` modifier, which got used in two ways:
1. A function marked `__intrinsic(glsl, ...)` had what I call a "target intrinsic" modifier, which specified how to lower it for a specific target (e.g., GLSL).
2. A function just marked `__intrinsic` was supposed to be a marker for "this function shouldn't be emitted in the output, because the implementation is expected to be provided"
The latter category of function should really be an `__intrinsic_op`, so I translated all those uses. I added a tiny bit of sugar so that `__intrinsic_op` without an explicit opcode will look up an opcode based on the name of the function being called, so that an operation like `sin` can automatically be plumbed through to an equivalent IR op. (The first category is a stopgap for the AST-based cross-compilation, and will hopefully be replaced by something better as we get the IR-based path working).
Getting the switch from `__intrinsic` to `__intrinsic_op` working required shuffling around some code in `emit.cpp` that handles looking up those modifiers and emitting builtin operations appropriately during cross-compilation.
Depending on where we go with things, a possible extension of this approach is to allow multiple operands to `__intrinsic_op` so that the first specifies the opcode, and then the rest are literal arguments to specify "sub-ops." This could help us handle stuff like texture-fetch operations without an explosion in the number of opcodes. I still need to think about whether this is a good idea or not.
|
|
This gets us far enough that we can convert a single test case to use the IR, under the new `-use-ir` flag.
Getting this merged into mainline will at least ensure that we keep the IR path working in a minimal fashion, even when we have to add functionality the existing AST-based path
There is definitely some clutter here from keeping both IR-based and AST-based translation around, but I don't want to have a long-lived branch for the IR that gets further and further away from the `master` branch that is actually getting used and tested.
Summary of changes:
- Add pointer types and basic `load` operation to be able to handle variable declarations
- Add basic `call` instruction type
- Add simple address math for field reference in l-value
- Always add IR for referenced decls to global scope
- Add notion of "intrinsic" type modifier, which maps a type declaration directly to an IR opcode (plus optional literal operands to handle things like texture/sampler flavor)
- Improve printing of IR instructions, types, operands
- Add constant-buffer type to IR
- Allow any instruction to be detected as "should be folded into use sites" and use this to tag things of constant-buffer type
- Also add logic for implicit base on member expressions, to handle references to `cbuffer` members
- Add connection back to original decl to IR variables (including global shader parameters...)
- Use reflection name instead of true name when emitting HLSL from IR (so that we can match HLSL output)
- Make IR include decorations for type layout
- Re-use existing emit logic for HLSL semantics to output `register` semantics for IR-based code
- Make IR-based codegen be an option we can enable from the command line
- It still isn't on by default (it can barely manage a trivial shader), but it seems better to enable it always instead of putting it under an `#ifdef`
- Fix up how we check for intrinsic operations suring AST-based cross compilation so that adding new intrinsic ops for the IR won't break codegen.
|
|
Previously, a `StructType` was an ordinary instruction that took a variable number of types are operands, representing the types of fields.
This ends up being inconvenient for a few reasons:
- To add decorations to the fields, you'd end up having to decorate the struct type instead (SPIR-V has this problem)
- You need to compute field indices during lowering, when you might prefer to defer that until later
- The get/set field operations now need an index, which needs to be an explicit operand, which means a magic numeric literal floating around to represent the index
The new approach fixes for the first two of these, and at least makes the last one a bit nicer.
A `StructDecl` is now a parent instruction, and its sub-instructions represent the fields of the type - each field is an explicit instruction of type `StructField`.
The operation to extract a field takes a direct reference the struct field, so everything is quite explicit.
|
|
- Change IR instructions to just hold an integer opcode instead of a pointer to the "info" structure
- Externalize definition of IR instructions to a header file, and use the "X macro" approach to allow generating different definitions
- Add notion of function types to the IR, so that we can easily query the result type of a function
- Add some convenience accesors to allow walking the IR in a strongly-typed manner (e.g., iterate over the parameters of a function)
- TODO: these should really be changed to assert the type of things, as least in debug builds
- Add very basic logic to `emit.cpp` so that it can walk the generated IR and start printing it back as HLSL
- This isn't meant to be usable as-is, but it is a step toward where we need to go
|
