| Age | Commit message (Collapse) | Author |
|---|
|
correctly mangled function.
|
|
|
|
|
|
|
|
|
|
There were some cases where we would try to emit an `ErrorType` to the output HLSL, which obviously isn't allowed. This change tries to avoid emitting error types, and instead use the original expression when it is available.
I tried adding a test case for the change, but I can't convince Slang to output matching line numbers for the error message; it seems like more work is needed on that front.
|
|
This is the first step towards supporting traditional object-oriented method definitions; the second step will be to allow `this` expressions to be implicit.
- Add a test case using explicit `this`, and expected output
- Update parsing logic for expressions so that it handled identifiers similarly to the declaration and statement logic: first try to parse using a syntax declaration looked up in the curent scope, and otherwise fall back to the ordinary `VarExpr` case.
* As long as I'm making that change: switch `true` and `false` to be parsed via the callback mechanism rather than be special-cased.
* This change will also help out if we ever wanted to add `super`/`base` expressions, `new`, `sizeof`/`alignof` or any other expression keywords.
- Add a `ThisExpr` node and register a parser callback for it.
- Add semantic checks for `ThisExpr`: basically just look upwards through scopes until we find either an aggregate type declaration or an `extension` declaration, and then use that as the type of the expression.
- TODO: eventually we need to guard against a `this` expression inside of a `static` member.
- The IR generation logic already handled creation of `this` parameters in function signatures; the missing piece was to register the appropriate parameter in the context, so that we can use it as the lowering of a `this` expression.
|
|
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.
|
|
|
|
inputs for running test shaders with arbitrary parameter definitions.
This commit contains the parser of the resource input definition.
|
|
* Fix up emission of shader parameter semantics when using IR
- Make sure to propagate entry point parameter layouts down to IR parameters when doing the initial cloning to form target-specific IR
- When layout information is present on an IR node, prefer to use that over the original high-level declaration for outputting semantics in final HLSL
- Fix up test runner to generate `.actual` files when running compute tests, in cases where the `render-test` application errors out (e.g., because of a Slang compilation error)
- Add a first test of generics functionality, to show that they generate valid code through the IR
- Right now this test is *not* using any "interesting" operations on the type parameter, so this is not a test that can confirm that interface constraints work
* fixup: skip compute tests when running on Linux
|
|
There was already a pass in place that transformed parameters and results of an entry-point function into global variables for GLSL, but this pass would just turn a `struct`-type parameter into a `struct`-type global, which has two problems:
- The standard GLSL language doesn't seem to allow `struct` types as vertex shader inputs or fragment shader outputs.
- If there are any members in such a `struct` that represent "system value" inputs or outputs, then these would need to be transformed into the equivalent `gl_*` variables.
This change adds a more complete scalarization process that applies to inputs/outputs during the legalization pass. In order to support this there is a little bit of a data strcuture for abstracting over tuples of values (this same idiom is used in a few other places, so perhaps the implementation could be done once and shared?).
System values are current handled in a painfully ad hoc (and incomplete) fashion during code emit. We need to come up with a better solution for mapping HLSL `SV_*` semantics over to `gl_*` variables.
In some cases this mapping might introduce more code than we can easily deal with during emit time, so it probably needs to be handled back at the IR level.
This implementation has many gaps, but it appears to be enough to get teh `render/cross-compile-entry-point` test working with IR-based cross-compilation.
|
|
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
|
|
The big addition here is that the Slang "bytecode" is no longer treated as just a "code generation target" (`CodeGenTarget`) akin to DX bytecode (DXBC) or SPIR-V, but instead is a `ContainerFormat` that can be used to emit all the results of a compile request (well, currently just the IR-as-BC, but the intention is there).
Getting to this goal involved some prior checkins that eliminated bogus "targets" that weren't really akin to SPIR-V or DXBC: `-target slang-ir-asm` and `-target reflection-json`. Those targets were really in place to support testing, and so they've been made more explicit testing/debug options.
This change eliminates `-target slang-ir` and instead tries to allow the user to specify `-o foo.slang-module` as an output file name, that indicates the intention to output a "container" file that will wrap up all the generated code.
I've also gone ahead and generalized the existing `-target` option so that we are actually building up a *list* of code generation targets. This is largely just a cleanup, since it forces code to be more aware of when it is doing something target-specific vs. target independent. For example, reflection layout information lives on a requested target, and not on the compile request as a whole, and similarly output code is per-target, per-entry-point.
As a cleanup, I eliminated support for per-translation-unit output. This was vestigial code from back when I used to try and do HLSL generation for a whole translation unit instead of per-entry-point (which turned out to be a lot of complexity for little gain), and it was only being used in the `hello` example and the `render-test` test fixture - in both cases fixing it up was easy enough. I've stubbed out the old `spGetTranslationUnitSource` API, but haven't removed it yet.
|
|
* Get rid of the `-slang-ir-asm` target
This is really only useful for debugging, so I've replaced the functionality with a `-dump-ir` command line option (which dump's the IR for an entry point before doing codegen).
* fixup: use HLSL target, not DXBC, so test can run on Linux
|
|
* Bug fix: emit logic for `do` loops
This case was never tested, and I was outputting some garbage characters. This comit fixes the codegen and adds a test case.
* Bug fix: make sure to pass through `[allow_uav_condition]`
This also fixes the standard library definition of `IncrementCounter()` so that it returns a `uint` instead of `void`.
|
|
Given an input declaration like:
cbuffer C
{
int a = -1;
}
Slang was automatically generating a `packoffset` semantic to place the member manually, but was emitting it *after* the initializer of the original declaration:
cbuffer C : register(b)
{
int a = -1 : packoffset(c0);
}
That syntax is invalid, of course, and we actually want:
cbuffer C : register(b)
{
int a : packoffset(c0) = -1;
}
This wasn't spotted earlier because putting initializers on a `cbuffer` member is not commonly done, since it requires reading those values via the reflection API. Slang's reflection API currently provides no way to access default values like this, so they aren't of much use yet. Still, it is better to emit correct syntax even in cases like this one.
|
|
When outputting GLSL from a Slang or HLSL entry point, we need to translate any parameters or results of an entry-point function into global declarations of `in` or `out` parameters, as needed by GLSL.
This change adds these transformations at the IR level, so that they don't need to complicate the emit logic.
More detailed changes:
- Make `render0` test use IR. It passes out of the box.
- Fix test runner to not always dump diffs on failures
I accidentally initialized an option to `true` instead of `false` when working on debugging the Travis CI failures.
- Special-case output for component-wise multiplication to handle GLSL `matrixCompMul()`
- Handle GLSL vs. HLSL output for calls to `mul()`
- Output proper `layout(std140)` on GLSL constant buffer declarations
- Require appropriate GLSL extension when emitting explicit `layout(offset = ...)` on constant buffer members
- TODO: Need to avoid requiring this extension in cases where the offsets are what would be computed anyway.
Realistically, should probably be emitting code with explicit padding, etc. to guarantee layouts.
- Add an IR-based pass to translate entry point functions by eliminating their input/output parameters and replacing them with global variables.
- Demangle names when calling target intrinsics
The lowering to the IR will turn a call like `sin(foo)` into a call to a function declaration with a mangled name like `_S3sin...`. This works fine when the user is calling their own functions, since the name mangling will apply to both the definition and use sites, but for builtin functions it obviously isn't what we want.
This change makes it so that we demangle the name of an instrinsic function just enough so that we can extract the original simple name, and make a call using that.
These changes do nor provide 100% of what we need when translating to GLSL, so the `cross-compile-entry-point` test *still* hasn't been flipped over to use the IR (even though that is the test case I've been using to develop these changes).
|
|
The main change I was working on here was to start having more of the builtin functions (in this case, `cos`, `sin`, and `saturate`) just lower to the IR as calls to builtin functions (with declarations but no definition), rather than expect/require them to map to individual IR opcodes in every case.
The main change there was the removal of some `intrinsic_op` modifiers in the stdlib. This then requires the `isTargetInstrinsic` logic in IR-based code emit to avoid emitting declarations for these intrinsics.
The corresponding logic for emitting *calls* to these intrinsics is currently being skipped.
Along the way, a variety of fixups were added:
- In order to support lowering to GLSL, we need to handle cases where a variable/function name uses a GLSL reserved word. The right long-term fix there is to always use generated or mangled names, but for now I'm hacking it by adding a `_s` prefix to all names during IR-based emit.
- This needs a flag to disable it, since some of our tests currently rely on checking binding information from generated HLSL/SPIR-V that will include these mangled/modified names.
- Emit matrix layout modifiers appropriately for GLSL
- Specialize IR parameter-block emission between GLSL and HLSL
- Fix up argument count/index logic for a couple of opcodes that weren't fixed when removing the types from the explicit operand list
- Fix up IR generation for calls to declarations with generic arguments. We were briefly adding the generic args to the ordinary argument list, which added complexity in several places. We now rely on the declaration-reference nodes in the IR to carry that extra info.
- TODO: We actually need to make sure that this is the case, since we don't currently correctly generated specialized decl-refs when building IR for function calls
The main test that would have been affected by this is `cross-compile-entry-point`, but I was not able to get that working fully with the IR. The main problem in this case was that when emitting GLSL we will need to perform certain required transformations on the IR to get legal code for GLSL. Notably:
- We need to hoist entry-point parameters away from being function parameters, and make them be global variables. This is currently being hand-waved during the emit logic, but it seems way better to have it all get cleaned up in the IR first.
- We need to scalarize entry-point parameters, because structure input/output is not supported as vertex input or fragment output (and it may be best to always scalarize anyway, to match HLSL semantics). (Note: "scalarize" here means to bust up structures, but not matrices/vectors)
|
|
* 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
|
|
* First attempt at a Linux build
- Fix up places where C++ idioms were written assuming lenient behavior of Microsoft's compiler
- Add a few more alternatives for platform-specific behavior where Windows was the only platform accounted for.
- Add a basic Makefile that can at least invoke our build, even if it isn't going good dependency tracking, etc.
- Build `libslang.so` and `slangc` that depends on it, using a relative `RPATH` to make the binary portable (I hope)
- Add an initial `.travis.yml` to see if we can trigger their build process.
* Fixup: const bug in `List::Sort`
I'm not clear why this gets picked up by the gcc *and* clang that Travis uses, but not the (newer) gcc I'm using on Ubuntu here, but I'm hoping it is just some missing `const` qualifiers.
* Fixup: reorder specialization of "class info"
Clang complains about things being specialized after being instantiated (implicilty), and I hope it is just the fact that I generate the class info for the roots of the hierarchy after the other cases. We'll see.
* Fixup: add `platform.cpp` to unified/lumped build
* Fixup: Windows uses `FreeLibrary`
and not `UnloadLibrary`
* Fixup: fix Windows project file to include new source file
This obviously points to the fact that we are going to need to be generating these files sooner or later.
|
|
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.
|
|
At a high level, this commit adds two things:
1. A "bytecode" format for serializing Slang IR instructions and related structure (functions, "registers")
2. A virtual machine that can load and then execute code in that bytecode format.
The reason for kicking off this work right now is that we *need* a way to run tests on Slang code generation that doesn't rely on having a GPU present (given that our CI runs on VM instances without GPUs), nor on textual comparison to the output of other compilers. With these features I've implemented a slapdash `slang-eval-test` test fixture that can run a (trivial) compute shader to very our compilation flow through to bytecode.
Some key design constraints/challenges:
- The bytecode format should be "position independent" so that a user can just load a blob of data and then inspect it without having to deserialize into another format, allocate memory, etc. Eventually the bytecode format might be a replacement for out current reflection API (we used to base reflection off a similar format, but the cost/benefit wasn't there at the time and we switched to just using the AST).
- The VM should be able to execute bytecode functions without doing any per-operation translation, JIT, etc. (translation of more coarse-grained symbols is okay). For now the VM is just being used to run tests, but eventually I'd like it to be viable for:
- Running Slang-based code in the context of the compiler itself. This starts with stuff like constant-folding in the front-end, but could expand to more general metaprogramming features.
- Running Slang-based ocde within a runtime application (e.g., a game engine) that wants to be able to run things like "parameter shader" code, or even just evaluate compute-like code on CPU (e.g., when supporting particles on both CPU and GPU).
- Finally, the bytecode format should ideally be able to round-trip back to the IR without unacceptable loss of information. This requirement and the previous one play off of each other, because things like a traditional SSA phi operation is ugly when you have to actually *execute* it. This doesn't matter right now when we don't have SSA yet, but it might be part of the decision-making here.
The actual implementation is centralized in `bytecode.{h,cpp}` and `vm.{h.cpp}`.
Big picture notes:
- The space of opcodes is shared between IR and bytecode (BC), with the hope that this makes translation of operations between the two easy.
- The actual bytecode instruction stream relies on a variable-length encoding for integer values, including opcodes and operand numbers, so that the common case is single-byte encoding.
- In the long term I intend to have a rule that if you use a single-byte encoding for an opcode, then all operands are required to use single-byte encodings too. Operations that need multi-byte operands would then be forced to use a multi-byte encoding of the op, and would be sent down a slower path in the interpeter.
- The "bytecode"'s outer structure is based on ordinary data structures linked with pointers, but they are "relative pointers" so the actual structure is position-independent.
- There are two main kinds of operands: registers and "constants." An operand is a signed integer where non-negatie values indicate registers (with `index == operandVal`) and negative values indicate constants (with `index == ~operandVal`).
- Registers are stored in the "stack frame" for a VM function call, and each has a fixed offset based on the size of the type and those that come before it. Conceptually, registers are allowed to overlap if they aren't live at the same time, and we manage this with a simple stack model: every register is supposed to identify the register that comes directly before it (this isn't implemented yet).
- "Constants" are more realistically a representation of "captured" values, but they are currently also how constants come in. Basically we can use a compact range of indices in the bytecode for a function, and each of these indices indirectly refers to some value in the next outer scope.
- The actual encoding of bytecode instructions right now is largely ad-hoc and very wasteful (we encode the type on everything, and we also encode everything as if it had varargs).
- In some cases, an instruction needs to know the types of the values involved (e.g., because it needs to load an array element, which means copying a number of bytes based on the size). The way the VM works we have types attached to our registers, so we currently get sneaky and look at those types in some ops. Longer term is makes sense to encode the required type info directly in the BC.
- There's a whole lot of hand-waving going on with how the actual top-level bytecode module gets loaded, because of the way we currently treat the top-level module as an instruction stream in the IR. This means that we try to represent the loaded module as a "stack frame" for a call to the module as a function, but that approach as serious problems, and isn't realistically what we want to do.
|
|
* 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
|
|
I had expected this to be the first case where control-flow instructions were needed, but it turns out that we aren't testing that entry point right now.
The real work/fix here ended up being handling of the `row_major` layout qualifier on a matrix inside a `cbuffer`. The existing AST-based code was passing it through easily (although I don't believe it was handling the layout rules right).
Getting it working in the IR involved beefing up the type-layout behavior so that it can handle explicit layout qualifiers (at least for matrix layout) which should also improve the layout computation for non-square matrices with nonstandard layout in the AST-based path.
There are still some annoying things left to do:
- The `row_major` and `column_major` layout qualifiers in HLSL/GLSL mean different things (well, they mean the reverse of one another) so I need to validate that the GLSL case is working remotely correctly.
- The layout logic isn't handling other explicit-layout cases as supported by GLSL (but of course GLSL is a far lower priority than HLSL/Slang)
- There is currently no way to pass in an explicit matrix layout flag to Slang to control the default behavior.
- Any client who was using Slang for HLSL pass-through and then applying a non-default flag on their HLSL->* compilation will get nasty unexpected behavior when the IR goes live - and they are already dealing with nasty behavior around non-square matrices not matching layout between Slang and the downstream.
- The logic that gleans layout modes from a variable declaration is currently only being applied for fields of structure types (which applies to `cbuffer` declarations as well), and not to global-scope uniform variables.
- We need test cases for all of this.
|
|
- 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, there were a variety of rules in `check.cpp` to pick the conversion cost for various cases involving scalar, vector, and matrix types.
- The main problem of the previous approach is that any lowering pass would need to convert an arbitrary "type cast" node into the right low-level operation(s).
- The new approach is that a type conversion (implicit or explicit) always resolves as a call to a constructor/initializer for the destination type. This means that the existing rules around marking operations as builtins should work for lowering.
- The support this, the checking logic needs to perform lookup of intializers/constructors when asked to perform conversion between types. It does this by re-using the existing logic for lookup and overload resolution if/when a type was applied in an ordinary context.
- Next, we define a modifier that can be attached to constructors/initializers to mark them as suitable for implicit conversion, and associate them with the correct cost to be used when doing overload comparisons.
- We add the modifier to all the scalar-to-scalar cases in the stdlib, using the logic that previously existed in semantic checking.
- Next we add cases for general vector-to-scalar conversions that also convert type, using the same cost computation as above.
- This probably misses various cases, but at this point they can hopefully be added just in the stdlib.
- One gotcha here is that in lowering, we need to make sure to lower any kind of call expression to another call expression of the same AST node class, so that we don't lose information on what casts were implicit/hidden in teh source-to-source case.
Two notes for potential longer-term changes:
1. There is still some duplication between the type conversion declarations here and the "join" logic for types used for generic arguments. Ideally we'd eventually clean up the "join" logic to be based on convertability, but that isn't a high priority right now, as long as joins continue to pick the right type.
2. It is a bit gross to have to declare all the N^2 conversions for vector/matrix types to duplicate the cases for scalars. For the simple scalar-to-vector case, we might try to support multiple conversion "steps" where both a scalar-to-scalar and a scalar-to-vector step can be allowed (this could be tagged on the modifiers already introduced). That simple option doesn't scale to vector-to-vector element type conversions, though, where you'd really want to make it a generic with a constraint like:
vector<T,N> init<U>(vector<U,N> value) where T : ConvertibleFrom<U>;
Here the `ConvertibleFrom<U>` interface expresses the fact that a conforming type has an initializer that takes a `U`. What doesn't appear in this context is any notion of conversion costs. We'd need some kind of system for computing the conversion cost of the vector conversion from the cost of the `T` to `U` converion.
|
|
The terminology here is similar to SPIR-V. For right now the only decoration exposed is a fairly brute-force one that just points back to a high-level declaration so that we can look up info on it that might affect how we print output.
|
|
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
|
|
- Make all instructions store their argument count for now, so we can iterate over them easily.
- Longer term we might try to optimize for space because the common case is that the operand count is known, but keeping it simpler seems better for now
- Split apart the creation of an instruction from adding it to a parent
- Use the above capability to make sure that we add a function to its parent *after* all the parameter/result type emission has occured.
- Perform simple value numbering for types during IR creation
- This logic also tries to pick a good parent for any type instructions, so that types don't get created local to a function unless they really need to
- Create all constants at global scope, and re-use when values are identical
|
|
- The changes introduced a new path where we don't even go through the current "lowering" (really an AST-to-AST legalization pass), but this exposed a few issues I didn't anticipate:
- First, we needed to make sure to pass in the computed layout information when emitting the original program (since the layout info is no longer automatically attached to AST nodes)
- Second, we needed to take the sample-rate input checks that were being done in lowering before, and move them to the emit logic (which is really ugly, but I don't see a way around it for GLSL).
|
|
With this change, basic generation of IR works for a trivial shader, and there is some basic support for dumping the generated IR in an assembly-like format.
As with the other IR change, the use of the IR is statically disabled for now, so that existing users won't be affected.
|
|
Fixes #23
Up to this point, the compiler has used the ordinary `String` type to represent declaration names, which means a bunch of lookup structures throughout the compiler were string-to-whatever maps, which can reduce efficiency.
It also means that things like the `Token` type end up carying a `String` by value and paying for things like reference-counting.
This change adds a `Name` type that is used to represent names of variables, types, macros, etc.
Names are cached and unique'd globally for a session, and the string-to-name mapping gets done during lexing.
From that point on, most mapping is from pointers, which should make all the various table lookups faster.
More importantly (possibly), this brings us one step closer to being able to pool-allocate the AST nodes.
|
|
This is in preparation for using `Name` as a type name.
|
|
Just like the previous change did for declaration keywords, this change uses the lexical environment to drive the lookup and dispatch of modifier parsing.
This allows us to easily add modifiers to Slang, even when they might conflict with identifiers used in user code (because the modifier names are no longer special keywords, but ordinary identifiers).
There was already some support for ideas like this with `__modifier` declarations (`ModifierDecl`) used to introduce some GLSL-specific keywords (so that they wouldn't pollute the namespace of HLSL files).
The new approach changes these to be actual `syntax` declarations (`SyntaxDecl`) with the same representation as those used to introduce declaration keywords.
Because many modifiers just introduce a single keyword that maps to a simple AST node (no further tokens/data), I modified the handling of syntax declarations so that they can take a user-data parameter, and this allows the common case ("just create an AST node of this type...") to be handled with minimal complications.
This also adds in a general-purpose string-based lookup path for AST node classes, that should support programmatic creation in more cases.
Statements are now the main case of keywords that need to be made table driven.
|
|
The existing parser code was doing string-based matching on the lookahead token to figure out how to parse a declaration, e.g.:
```
if(lookAhead == "struct") { /* do struct thing */ }
else if(lookAhead == "interface") { /* do interface thing * }
...
```
That approach has some annoying down-sides:
- It is slower than it needs to be
- It is annoying to deal with cases where the available declaration keywords might differ by language
- Most importantly, it is not possible for us to introduce "extended" keywords that the user can make use of, but which can be ignored by the user and treated as an ordinary identifier.
That last part is important. Suppose the user wanted to have a local variable named `import`, but we also had a Slang extension that added an `import` keyword. Then a line of code like `import += 1` would lead to a failure because we'd try to parse an import declaration, even when it is obvious that the user meant their local variable. This would mean that Slang can't parse existing user code that might clash with syntax extensions. This issue is the reason why we currently have keywords like `__import`.
A traditional solution in a compiler is to map keywords to distinct token codes as part of lexing, which eliminates the first conern (performance) because now we can dispatch with `switch`. It can also aleviate the second concern if we add/remove names from the string->code mapping based on language (the rest of the parsing logic doesn't have to know about keywords being added/removed).
The solution we go for here is more aggressive.
Instead of mapping keyword names to special token codes during lexing, we instead introduce logical "syntax declarations" into the AST, which are looked up using the ordinary scoping rules of the language.
Depending on what code is imported into the scope where parsing is going on, different keywords may then be visible.
This solves our last concern, since a user-defined variable that just happens to use the same name as a keyword is now allowed to shadow the imported declaration for syntax (this is akin to, e.g., Scheme where there really aren't any "keywords").
This also opens the door to the possibility of eventually allowing user to define their own syntax (again, like Scheme).
For now I'm only using this for the declaration keywords.
With this change it should be pretty easy to also add statement keywords in the same fashion.
|
|
Fixes #24
So far the code has used a representation for source locations that is heavy-weight, but typical of research or hobby compilers: a `struct` type containing a line number and a (heap-allocated) string.
This is actually very convenient for debugging, but it means that any data structure that might contain a source location needs careful memory management (because of those strings) and has a tendency to bloat.
The new represnetation is that a source location is just a pointer-sized integer.
In the simplest mental model, you can think of this as just counting every byte of source text that is passed in, and using those to name locations.
Finding the path and line number that corresponds to a location involves a lookup step, but we can arrange to store all the files in an array sorted by their start locations, and do a binary search.
Finding line numbers inside a file is similarly fast (one you pay a one-time cost to build an array of starting offsets for lines).
More advanced compilers like clang actually go further and create a unique range of source locations to represent a file each time it gets included, so that they can track the include stack and reproduce it in diagnostic messages.
I'm not doing anything that clever here.
|
|
- `ExpressionSyntaxNode` becomes `Expr`
- `StatementSyntaxNode` becomes `Stmt`
- `StructSyntaxNode` becomes `StructDecl`
- `ProgramSyntaxNode` becomes `ModuleDecl`
- `ExpressionType` becomes `Type`
- Existing fields names `Type` become `type`
- There might be some collateral damage here if there were, e.g., `enum`s named `Type`, but I can live with that for now and fix those up as a I see them
|
|
There were two main places where global variables were used in the Slang implementation:
1. The "standard library" code was generated as a string at run-time, and stored in a global variable so that it could be amortized across compiles.
2. The representation of types uses some globals (well, class `static` members) to store common types (e.g., `void`) and to deal with memory lifetime for things like canonicalized types.
In each case the "simple" fix is to move the relevant state into the `Session` type that controlled their lifetime already (the `Session` destructor was already cleaning up these globals to avoid leaks).
For the standard library stuff this really was easy, but for the types it required threading through the `Session` a bit carefully.
One more case that I found: there was a function-`static` variable used to generate a unique ID for files output when dumping of intermediates is enabled (this is almost strictly a debugging option).
Rather than make this counter per-session (which would lead to different sessions on different threads clobbering the same few files), I went ahead and used an atomic in this case.
Note that the remaining case I had been worried about was any function-`static` counter that might be used in generating unique names.
It turns out that right now the parser doesn't use such a counter (even in cases where it probably should), and the lowering pass already uses a counter local to the pass (again, whether or not this is a good idea).
This change should be a major step toward allowing an application to use Slang in multiple threads, so long as each thread uses a distinct `SlangSession`. The case of using a single session across multiple threads is harder to support, and will require more careful implementation work.
|
|
- We use this to work around the fact that, e.g., `Texture2D.Load` doesn't take a sampler, but the equivalent GLSL operation `texelFetch` requires one
- Previously we tried to hide the sampler from the user, hoping that glslang would drop it and we could just ignore it, but that doesn't work
- For now we'll go ahead and explicitly show the sampler in the reflection info so that an app can react appropriately
- We also generate a unique binding for the sampler, instead of the old behavior that fixed it with `binding = 0`
- We still fix it with `set = 0`, so it might still surprise users
|
|
- API users can use this to get "clean" output to aid with debugging Slang issues
- Also changes the prefix on intermediate files that Slang dumps, to make them easier to ignore with a regexp
|
|
- The easy part here is treating `NV_` prefixed semantics as another case of "system-value" semantics
- Mapping the new semantics (`NV_X_RIGHT` and `NV_VIEWPORT_MASK`) to their GLSL equivalents is harder
- Instead of a single "right-eye vertex" output, GLSL defines an array of per-view positions
- Instead of a vector of masks, GLSL defines an array of per-view masks
- Another point here is that a lot of semantics that appear as `uint` in HLSL are `int` in GLSL, which can lead to conversion issues.
- The approach here is to have the lowering pass introduce a notion of assignment with "fixups," which will try to cast things as needed
- When assigning to a simple value with the "wrong" type, introduce a cast
- When assigning to an array from a vector, break out multiple assignments of individual vector/array elements
- In order to facilitate the above, I needed to add actual types to the magic expressions I introduce to represent GLSL builtin variables. These were taken by scanning the online documentation for GL, so they might not be perfect.
- Major issues with the approach in this change:
- No attempt is being made here to check that the original declaration used a type appropriate to the semantic. The assumption is that this logic only ever triggers for Slang entry points, or GLSL entry points using a Slang `struct` type for input/output (and for right now Slang code is only ever written by "understanding" developers)
- In the case of a Slang entry point, we always copy varying parameters in/out around the call to `main_`, so this approach should handle calls to functions with `out` or `in out` parameters okay, but it is *not* robust to cases where we don't want to copy in all the entry point parameters first thing (e.g., a GS), so that will have to change
- In the GLSL case (or if we revise the approach to Slang entry points), there is going to be a problem if these converted varying parameters are ever passed as arguments to `out` or `in out` parameters. In these cases we need to do more sleight-of-hand to reify a temporary variable and do the necessary copy-in/copy-out. Being able to do that logic relies on having correct information about callees, which requires having robust semantic analysis of the function body. There is only so much we can do...
- A better long-term approach would not rely on an ad-hoc "fixup" conversion during assignment, but would instead implement the GLSL builtin variables as, effectively, global "property" declarations that have both `get` and `set` accessors, and then tunnel a reference to such a property down through lowering, where it can lower to uses of the "getter" or "setter" as appropriate in context (and the result type of the getter/setter can be what we'd want/expect).
|
|
The change is mostly about trying to make sure the compiler "fails safe" when it encounters an internal assumption that isn't met.
Most internal errors will now throw exceptions (yes, exceptions are evil, but this will work for now), and these get caught in `spCompile` so that they don't propagate to the user (they just see a message that compilation aborted due to an internal error).
Subsequent changes are going to need to work on diagnosing as many of these situations as possible, so that users can at least know what construct in their code was unexpected or unhandled by the compiler.
|
|
Fixes #122
- In cases with an explicit mip level being specified, there was a mistake in how the argument for setting the mip level in the GLSL code was constructed that led to a parse error in GLSL
- Also, that argument is a `uint` in HLSL and an `int` in GLSL, so an explicit cast was needed
- The GLSL functions here seem to require a newer GLSL (at least higher than `420`), so I had to add in a capability for builtins to specify a required GLSL version. For now I made these ones require `450`.
- Added a test case to confirm that our lowering works (for some definition of "works")
|
|
When lowering `buf[i]` to `texelFetch(..., i)` we need to deal with the case where the type of `b` might be `Buffer<float>` in which case we want to add a `.x` swizzle to the end of the fetch.
|
|
GLSL technically supports varying (`in`, `out`) parameters of `struct` type, but there are some annoying constraints (not allowed for VS input), and it doesn't work with how an HLSL user would usually put "system-value" inputs/outputs into a `struct` together with ordinary inputs/outputs.
To work around this, this change adds support for using an imported Slang `struct` type for an `in` or `out` parameter, in which case it will (1) be scalarized and (2) will have system-value semantics mapped appropriately, just as for an entry-point parameter when cross-compiling an HLSL-style `main()`.
Changes:
- Add a notion of a `VaryingTupleExpr` and `VaryingTupleVarDecl`, similar to those for the resources-in-structs case
- Trigger use of these when we have a global-scope varying in/out using an imported `struct` type
- Also use these in the cross-compilation case for ordinary varying input/output (since this approach seems like it should be more general, and can hopefully handle stuff like GS input/output some day)
- When generating parameter binding information, special case global-scope input/output, and treat it the same as entry-point-parameter input/output
- Revamp how used resource ranges are computed so that we can eventually make this specific to an entry point
- Actually implement first signs of life for `maybeMoveTemp` so that assignments to the tuple-ified outputs will work better
- Add first test case that actually seems to work
- Add diagnostics for conflicting explicit bindings on a parameter
- Add diagnostic for different parameters with overlapping bindings
- Make global-scope varying input/output use a tracking data structure specific to the translation unit for computing locations (so that they are independent of other TUs)
|
