FXC fails to compile code that assigns to dynamically-indexed fixed-size
arrays in structs on internal shader variables with:
error X3500: array reference cannot be used as an l-value; not natively
addressable
This CL detects this case, and transforms such assignments into copying
out the array to a local variable, assigning to that local, and then
copying the array back.
Also manually regenerate SKIPs for HLSL/FXC after this change, which
fixes 30 tests. Also exposes some "compilation aborted unexpectedly" now
that "array reference cannot be used as an l-value" has been fixed. For
tests that fail for both DXC and FXC, updating SKIPs to the DXC one to
help distinguish actual FXC bugs from valid errors.
Bug: tint:998
Bug: tint:1206
Change-Id: I09204d8d81ab27d1c257538ad702414ccc386543
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/71620
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Any struct which is used as the store type of a buffer variable needs
to have a block decoration. If that struct is nested inside an array
or another struct, we wrap it inside another struct first.
This removes the SPIR-V backend's reliance on the [[block]] attribute,
which will soon be deprecated and removed.
Bug: tint:1324
Change-Id: Ib6ad54f24a3e4a090da9faeed699f266abcb66ba
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/72082
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
This transform does not use the top-level ctx.Clone() method, so we
need to manually copy the applied transforms over to the destination
program.
This fixes the GLSL backend, where this transform was dropping the
applied transforms and breaking a required transform dependency later
in the chain.
Change-Id: I3db688fe35dfbe67468b9941f739cb2ecf3a843b
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/72220
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Migrate some of the validation logic over to use the results of behavior
analysis.
The most significant changes are:
* Unreachable-statements now consider merge-points of control flow. For
example, if all branches of a if-statement or switch-statement either
return or discard, the next statement will be considered unreachable.
* Unreachable statements are no longer an error, but a warning. See
https://github.com/gpuweb/gpuweb/issues/2378.
* Statements that follow a loops that does not break, or have a
conditional will now be considered unreachable.
* Unreachable statements produced by the SPIR-V reader are now removed
using the new RemoveUnreachableStatements transform.
Some other new changes include additional validation for the continuing
block for for-loops, to match the rules of a loop continuing block.
The new cases this validation is testing for are not expressible in
WGSL, but some transforms may produce complex continuing statements that
might violate these rules. All the writers are able to decay these
complex for-loop continuing statements to regular loops.
Bug: tint:1302
Change-Id: I0d8a48c73d5d5c30a1cddf92cc3383a692a58e61
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/71500
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: David Neto <dneto@google.com>
Add transform::Unshadow to renamed shadowed symbols. Required by a
number of other transforms.
Replace Resolver symbol resolution with dep-graph.
The dependency graph now performs full symbol resolution before the
regular resolver pass.
Make use of this instead of duplicating the effort.
Simplfies code, and actually performs variable shadowing consistently.
Fixed: tint:819
Bug: tint:1266
Change-Id: I595d1812aebe1d79d2d32e724ff90de36e74cf4b
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/70523
Reviewed-by: David Neto <dneto@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
They are always used together, and can actually be simplified when
combined.
This transform cannot currently deal with shadowing, and will need
need to depend on a new 'unshadow' transform.
Fixes a long-standing bug where we'd get an ICE if we attempt to inline
a pointer let declaration with side-effects in a for-loop initializer.
Fixed: tint:1321
Bug: tint:819
Change-Id: I236fed688e33a4996e47310b5ece44c991b5249f
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/70661
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Performs a module-scope (global) declaration dependency analysis, so
that out-of-order global declarations can be re-ordered into dependency
order for consumption by the resolver.
The WGSL working group are currently debating whether out-of-order
declarations should be included in WebGPU V1, so this implementation
currently errors if module-scope declarations are declared out-of-order,
and the resolver does not currently use this sorted global list.
The analysis does however provide significantly better error diagnostics
when cyclic dependencies are formed, and when globals are declared
out-of-order.
The DependencyGraph also correctly now detects symbol collisions between
functions and types (tint:1308).
With this change, validation is duplicated between the DependencyGraph
and the Resolver. The now-unreachable validation will be removed from
the Resolver with a followup change.
Fixed: tint:1308
Bug: tint:1266
Change-Id: I809c23a069a86cf429f5ec8ef3ad9a98246766ab
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69381
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
ArrayLengthFromUniform is needed for correct bounds checks on
dynamic storage buffers on D3D12. The intrinsic GetDimensions does
not return the actual size of the buffer binding.
ArrayLengthFromUniform is updated to output the indices of the
uniform buffer that are statically used. This allows Dawn to minimize
the amount of data needed to upload into the uniform buffer.
These output indices are returned on the HLSL/MSL generator result.
ArrayLengthFromUniform is also updated to allow only some of the
arrayLength calls to be replaced with uniform buffer loads. For HLSL
output, the remaining arrayLength computations will continue to use
GetDimensions(). For MSL, it is invalid to not specify an index into
the uniform buffer for all storage buffers.
After Dawn is updated to use the array_length_from_uniform option in the
Metal backend, the buffer_size_ubo_index member for MSL output may be
removed.
Bug: dawn:429
Change-Id: I9da4ec4a20882e9f1bfa5bb026725d72529eff26
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69301
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Austin Eng <enga@chromium.org>
Implements MultiplanarExternalTextureTransform to allow transforming a
texture_external binding into two texture_2d<f32> bindings and a uniform
buffer binding. Transforms textureSampleLevel and textureLoad calls with
a texture_external parameter into custom functions that can handle both
single-plane RGBA or bi-planar YUV. Includes tests.
Bug: dawn:1082
Change-Id: Icb6d8b0f3773feca01c833171f07230c3531f3aa
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/68620
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Add a new 'Target' to the ast::CallExpression, which can be either an
Identifier or Type. The Identifier may resolve to a Type, if the Type is
a structure or alias.
The Resolver now resolves the CallExpression target to one of the
following sem::CallTargets:
* sem::Function
* sem::Intrinsic
* sem::TypeConstructor
* sem::TypeCast
This change will allow us to remove the type tracking logic from the WGSL
parser, which is required for out-of-order module scope declarations.
Bug: tint:888
Bug: tint:1266
Change-Id: I696f117115a50981fd5c102a0d7764641bb755dd
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/68525
Reviewed-by: David Neto <dneto@google.com>
Reviewed-by: James Price <jrprice@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Change all the texture types to the GLSL equivalents. Note that some
types don't actually exist in GLSL ES, e.g., 1D textures, but this
will be handled later.
Change all the texture functions from HLSL-style to GLSL (e.g.,
texture.Sample(...) -> texture(texture, ...). Note that depth
comparison functions are probably wrong.
Implement writeonly storage texture type and functions.
Samplers are skipped entirely in the GLSL backend, with the assumption
that they will already have been combined into GLSL-style combined
samplers and textures by the client code.
Move the SingleEntryPoint transform above the RemovePhonies pass. This
ensures that texture variables are not optimized out. (Otherwise some
tests produce valid but not very useful results.)
Add the builtin-functions to the GLSL keywords list for renaming.
They're not keywords, but they can't be identifiers either.
Bug: tint:1298, tint:1299
Change-Id: I86c4547fcdd1eba80be98f6c05b939f345fd4c3e
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69200
Commit-Queue: Stephen White <senorblanco@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Replace all calls to create<ast::TypeConstructorExpression>() with
Construct(). TypeConstructorExpression will be folded into
ast::CallExpression, but the Construct() call signature will remain
identical.
Bug: tint:888
Change-Id: Ifb28a90ccf5184c8090c2e32fa8c82f3996dfa33
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/69108
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Literals are now expressions, so in keeping with all the other
expression types, suffix the class name with Expression.
I'm not overly keen on requiring everything to have an Expression
suffix, but consistency is better than personal preference.
Note: this should have been part of 30848b6, but I managed to drop
this change instead of squashing it. Opps.
Bug: tint:888
Change-Id: Idfaf96abe165a6bf5028e60a160e7408aa2bf9db
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/68943
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Just make Literal an expression. The ScalarConstructorExpression
provides no real value, aside from having a ConstructorExpression base
class that's common between ScalarConstructorExpression and
TypeConstructorExpression. TypeConstructorExpression will be folded into
CallExpression, so this hierarchy will serve no purpose.
First step in resolving the parser ambiguity of type-constructors vs
type-casts vs function calls.
Bug: tint:888
Change-Id: I2585d5ddbf6c0619a8f24c503e61ebf27c182ebe
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/68524
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
The semantic nodes cannot be fully immutable, as they contain cyclic
references. Remove Resolver::CreateSemanticNodes(), and instead
construct and mutate the semantic nodes in the single traversal pass.
Give up on trying to maintain the 'authored' type names (aliased names).
These are a nightmare to maintain, and provided limited use.
Significantly simplfies the Resolver, and allows us to generate more
semantic to semantic references, reducing sem -> ast -> sem hops.
Note: This change introduces constant value propagation across constant
variables. This is unlocked by the earlier construction of the
sem::Variable.
Change-Id: I592092fdc47fe24d30e512952511c9ab7c16d7a1
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/68406
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
And remove a whole load of const_cast hackery.
Semantic nodes may contain internally mutable fields (although only ever modified during resolving), so these are always passed by `const` pointer.
While all AST nodes are internally immutable, we have decided that pointers to AST nodes should also be marked `const`, for consistency.
There's still a collection of const_cast calls in the Resolver. These will be fixed up in a later change.
Bug: tint:745
Change-Id: I046309b8e586772605fc0fe6b2d27f28806d40ef
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/66606
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
Methods and functions are `CamelCase()`
Public fields are `snake_case` with no trailing `_`
Private fields are `snake_case` with a trailing `_`
Remove pointless getters on fully immutable fields.
They provide no value, and just add `()` noise on use.
Remove unused methods.
Bug: tint:1231
Change-Id: If32efd039df48938efd5bc2186d51fe4853e9840
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/66600
Reviewed-by: David Neto <dneto@google.com>
Commit-Queue: Ben Clayton <bclayton@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
These are legacy methods that were written before the semantic type nodes.
These methods do not consider aliases, and any use of these is likely to be broken for aliases.
Fix up uses of these methods to use the semantic types instead.
Change-Id: Ia66749b279eddff655d3d755fef54a6263643e69
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/66601
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
These methods are going to be removed as they provide little benefit over the WGSL form, are a maintainance burden and they massively bloat our codebase.
This change introduces sem::CallTargetSignature, which can be used as a std::unordered_map key.
This is used in writer/spirv to replace a map that was keyed off ast::Function::type_name().
Bug: tint:1225
Change-Id: Ic220b3155011f21b14d49eecc8042001148e4ca5
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/66443
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: David Neto <dneto@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
This is a modified version of the HLSL writer.
Basic types, arrays, entry points, reserved keywords, uniforms,
builtin uniforms, structs, some builtin functions, zero initialization
are implemented. Textures, SSBOs and storage textures in particular are
unimplemented. All the unit tests "pass", but the output is not correct
in many cases.
triangle.wgsl outputs correct vertex and fragment shaders that pass
GLSL validation via glslang. compute_boids.wgsl outputs a valid but not
correct compute shader.
Change-Id: I96c7aaf60cf2d4237e45d732e5f51b345aea0552
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/57780
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
Reviewed-by: Ben Clayton <bclayton@google.com>
MSL has a limit on the number of threadgroup memory arguments, so use
a struct to support an arbitrary number of workgroup variables.
This commit introduces a `State` object to this transform, which is
used to track which structs have been cloned eagerly, in order to
avoid duplicating them.
Bug: tint:938
Change-Id: Ia467db186e176a08f160455eab5fd3b3662f56b8
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/65360
Auto-Submit: James Price <jrprice@google.com>
Kokoro: James Price <jrprice@google.com>
Commit-Queue: James Price <jrprice@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
