Pointers require decorating the variable with a * to read back the
value, which the code generator did not properly handle. There was a
special case to add the * but it only supported assignment into the
variable, not reading back. References require no special decoration.
This change fixes compile errors in Functions.sksl with the "bar"
function. (This test marks `x` as an inout but never actually mutates
it.) It also allows us to remove a special-case workaround for `frexp`,
an intrinsic function which uses a reference for its out-parameter.
Additionally, this CL adds a non-inlining copy of "OutParams.sksl" to
the Metal test directory, as most of our tests which use out-parameters
end up inlining all the code, which hides these sorts of bugs.
Change-Id: I31c4db04f6b512b4cd4fe65b3347b82bdbf039cd
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/341000
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, we would emit an invalid [[buffer(-1)]] annotation on the
block, causing the Metal compilation to fail.
Change-Id: I68b2439c05db3163686e84c5dcc9a5c43870ff67
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340761
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
It's not legal to use identifiers like "int" or "sampler" to name your
variables (or enums, or structs, etc.). SkSL will now report this as an
error instead of relying on the driver to catch this.
(Note that in some contexts, it might be legal by the spec to reuse a
name that you introduced yourself, depending on the scope. In practice,
this confuses Apple GLSL, so we shouldn't support it anyway.)
This caught several existing places in our code where we used the name
"sampler." These were never exposed to the driver (they were intrinsics
that we would replace during compilation) so they were harmless before.
Change-Id: Ia6dcfca8c500d02e1eb5f9427bed8727e114dfc2
Bug: skia:11036
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340758
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, our AST structures would include a "sizeCount" for arrays,
which indicated the number of AST nodes associated with array
dimensions. Since GLSL only supports a single array dimension, this
field has been replaced with "isArray," a boolean indicating whether we
have a single AST node for array size. This allowed many array-size
based looping constructs to be replaced with simpler non-looping
equivalents.
This change flushed out a few places where the parser was not actually
enforcing its promised maximum array-dimensionality.
Also found some duplicated code in variable-declaration parsing,
related to parsing array-sizes and initializer expressions. This has
been de-duplicated by using a lambda. (This change was likely why this
CL was not net-negative for LOC, but it's simpler and cleaner.)
Change-Id: I7abed732d3a296edf02c0ec9813fceb5aae4a9a0
Bug: skia:11026
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340656
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
GLSL only allows one-dimensional arrays. This CL lowers SkSL's array
dimensionality limit from eight to one, and fixes all the tests that
this breaks. The rest of the code still technically supports
arbitrarily-deep array dimensionality; there are many opportunities for
code cleanup and simplification in followup CLs.
Change-Id: I0fc31e4626649ec69d40c5f5597b3924de298df0
Bug: skia:11026
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340339
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This is illegal in older versions of GLSL and in Metal. We now fail at
SkSL compilation time and properly report the error.
Change-Id: I6ddaeabff5386a1ed6ca3eb8703a6035476ec77a
Bug: skia:11021
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339298
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This CL fixes cases where array dimensions could be placed on the type
instead of the variable (`float[2] x` instead of `float x[2]`). It also
reports errors in cases where arrays aren't syntactically valid in
Metal, rather than emitting unusable Metal code. (Some of these cases
are actually invalid GLSL as well! But those fixes are coming in
followup CLs.)
Change-Id: I22279127c8a9aa2f22bf5ea3d225e563c2e254f2
Bug: skia:10926, skia:10760, skia:10761
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340137
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
The proper approach for creating multi-dimensional array types is
complicated, so I added a function in SymbolTable which does it the
right way (addArrayDimensions). I found all the places in SkSL which
created arrays from base types and size arrays, and refactored them to
call addArrayDimensions instead of doing it manually.
I believe that this approach fixes a bunch of minor issues with multi-
dimensional array types; some are visible in the current codegen output,
and others are latent bugs. e.g. in some instances, a Variable's type()
was silently holding flipped array dimensions, but this never led to
a visible bug because we ended up using the VarDeclaration's baseType()
plus sizes() everywhere that the type was used. (In particular, this
caused debugging headaches in http://review.skia.org/340137 where I'd
use a Variable's type and suddenly its array dimensions would be wrong.)
Change-Id: Idd6a86aa5d1dce8918d02a53bcc2f7d7886e3ac5
Bug: skia:11016, skia:10924
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339860
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This reintroduces the flipped-array-dimensions bug in skia:10924. It
will be fixed in followup CLs.
Change-Id: I24ec687209b397f5fd0cf44194d0e21fe30dc32c
Bug: skia:10924
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339797
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The Metal return type from main() diverges from the SkSL source, so we
patch it in the Metal code generator. This CL improves the patching
process in multiple ways:
- A `return` statement from a fragment processor main() is rewritten to:
return *_out;
- A `return` statement from a vertex processor main() is rewritten to:
return (_out->sk_Position.y = -_out->sk_Position.y, *_out);
- We avoid emitting a duplicate `return *_out;` statement if we can
determine that main() already ends in a return statement. This is
harmless either way so it doesn't necessarily catch everything. (e.g.
it doesn't detect an if/else which returns at the end of both blocks.)
Also added a unit test which returns from the middle of a vertex shader,
since we didn't test this anywhere and we need to verify that
sk_Position.y will be negated. (This didn't work properly before.)
Change-Id: I14cf18375894fc712fa6c6466df3888ebaeba7c8
Bug: skia:10903
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339636
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, this would generate invalid code such as `[[user(locn-1)]]`.
We now generate a more-useful error at SkSL compilation time.
Change-Id: Ifbe335ec6d4abcbdfe89b892ba51063c94d22b11
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339397
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
GLSL only supports arrays of samplers in very limited ways; they aren't
supported at all by SkSL. We now detect arrays of opaque objects and
reject the code.
We have several paths through the IR generator that create and process
array types; the unit test covers global and local variables, and array
on the type versus array on the variable.
Change-Id: I5b45e88e31cf4005723c3bf35561622d65321f7b
Bug: skia:11008
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339317
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Just filling in a gap in our tests. The output is a little strange as it
exposes a missed opportunity to constant-fold array accesses, but it
seems fine otherwise.
Change-Id: I6df13e0f9a49455015ceb47d7802bb5e1bbdaa1a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339217
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Constructors such as `float[2](0, 0)` add a type to the symbol table;
this type needs to be copied into the new symbol table if the
constructor is cloned by the inliner.
Change-Id: Ifa8d2dec87103c6223ce493e2201a904c14c2137
Bug: oss-fuzz:28050
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339168
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
SPIR-V previously didn't know what to think when it encountered a Type
with a typeKind of kEnum, and would abort. These are now treated as
32-bit signed integers.
Metal previously emitted the SkSL enum typename, which is meaningless to
Metal since we do not emit the enum itself anywhere. Metal now emits
"int" for an enum-typed variable.
(GLSL already correctly emits "int" for enum types.)
Change-Id: I05975a2a399f9c4a22c00c90be0dccacd99d793b
Bug: skia:11003
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338856
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This CL addresses the root cause of the fuzzer issue, by checking for
LayoutIsSupported before getting the MemoryLayout of a type. However,
this array ought to be detected as an error everywhere, as samplers are
opaque types; at present, this code compiles without error in GLSL and
Metal. This is an issue for followup CLs.
GLSL's actual support for arrays of samplers is interesting and probably
too nuanced for us to try to emulate:
https://www.khronos.org/opengl/wiki/Data_Type_(GLSL)#Opaque_arrays
"Under GLSL version 3.30, Sampler arrays (the only opaque type 3.30
provides) can be declared, but they can only be accessed by compile-time
integral Constant Expressions. So you cannot loop over an array of
samplers, no matter what the array initializer, offset and comparison
expressions are.
Under GLSL 4.00 and above, array indices leading to an opaque value can
be accessed by non-compile-time constants, but these index values must
be dynamically uniform. The value of those indices must be the same
value, in the same execution order, regardless of any non-uniform
parameter values, for all shader invocations in the invocation group."
Change-Id: Ib382f5c3b563f996b3c8f1eb6b021b6d31fa9ce7
Bug: oss-fuzz:28107
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339159
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, GLSL and Metal code generators would emit a struct wherever
the type was first used in the code, regardless of where it was
originally defined or what scope the type needs to live in. This CL adds
a ProgramElement for struct definitions, so that structs will now appear
at the top-level as they were originally defined. In the case of Metal,
some special handling is also needed to handle the Globals struct
properly.
Not yet fully supported:
- No special handling for structs declared inside functions yet
- No support for structs in separate scopes with overlapping names
The severity of the remaining issues depends mostly on whether we want
to support structs inside functions in Runtime Effects.
Change-Id: Ia95d4529506cb3fa6da63f5cb548199a93e1c0c5
Bug: skia:10922, skia:10923, skia:10925, skia:10926
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338600
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This test verifies that dead-stripping works on both built-in and user
functions, if their function call is optimized away.
Change-Id: I3125a34640c69de43c383343cd00d97e5a32ac60
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338836
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Enums are an SkSL-only concept--when we output code, we emit plain
IntLiterals--so the fix is simply to ignore the Enum program element
when we encounter it. This is what GLSLCodeGen does as well.
Also added a unit test to confirm that enums work normally, and that
enums are subject to optimization and static-comparison checks just as
ints would be.
Change-Id: Ic4f8da7a27983add9eb41b936d46f6638d22bd4b
Bug: skia:11003
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338800
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
There were a surprisingly small number of dedicated SPIR-V tests.
SkSLSPIRVBadOffset was the only test that didn't already exist in the
golden outputs, although it actually contained two tests.
The SPIRVTest.cpp file has been converted to SPIRVTestbed.cpp, which can
be used for local debugging of SPIR-V issues via dm (like GLSLTestbed
and MetalTestbed).
Change-Id: I978d8a7cf5735af7f537113d2b9411ce42cfcf88
Bug: skia:10694
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338756
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
^^ is not an operator in Metal. != can be used for the same purpose.
Change-Id: If75b000076ebe0aa81d0ab354a8ae33e6ed52101
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/339156
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Change-Id: Ifc1f0921d983ee09d7bc2632aeca41689f1bf0c3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338603
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, we allowed unary minus on numbers and vectors (of any type).
Now, we allow them on numbers and vectors of numbers.
Also updated the Boolean arithmetic error test to cover scalars as well
as vectors.
Change-Id: Ie74d1f3bfc1e9353e04c6f8e468fa20e0cbba16f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338396
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
GLSL does not allow most binary operations on bvec types; we can now
detect these and properly flag them as errors.
Note that `determine_binary_type` was also refactored. It originally
started with an enormous omni-switch over every possible Token type,
used to set various bools describing the type of binary expression at
hand. Instead of one big switch, this has been refactored into several
small switches in standalone functions that simply switch on the op and
immediately return true or false. Conceptually this seems like more
work (checking the op multiple times), but these tiny switches actually
boil down to little branchless shift-and-mask functions, so in practice
they should be quite efficient compared to the original omni-switch.
Change-Id: I81b473d98c65da1edd136f35fc8f656261f8930d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338346
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This is very unlikely to occur in real-world code, as it's somewhat
nonsense to use the comma operator in this way. However, it's better to
fail cleanly than to assert.
Change-Id: I76481cd8a993cb1a798ee16956400a512efd4c15
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337636
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Fiddled with the logic a bit so that when we're in unit test mode, the
output still includes all of the SPIR-V (as well as the validation error
message), so that tracking them down is easier.
Bug: skia:10694
Change-Id: I15e7777af3d268a5952765dbe5d63612cad0ac07
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338320
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Fix code generation for Metal and Vulkan with geometric
intrinsics that have scalar versions in GLSL/SkSL, but no
native support in MSL/SPIR-V.
Change-Id: Id4538a00172e0d233ad9d5ed8d33db6436b83208
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338276
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Previously, we assumed that if a vector in `is_constant` was not made of
floats, it must be made of integers. This ignores that boolean vectors
also exist. The original code would abort when `getIVecComponent` was
called on a bool vector.
There is another bug here--arithmetic operators on bool types should be
disallowed entirely. That will be addressed in later CLs.
Change-Id: I78781d839abde9376917fd92f2fe6311a1a58b02
Bug: oss-fuzz:27808
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/338055
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This reverts commit e81fb87bb4.
Reason for revert: checking results with less-aggressive inliner
Original change's description:
> Revert "Simplify _blend_set_color_saturation, removing an instruction."
>
> This reverts commit ed289e777c.
>
> Reason for revert: causing strange artifacts, only on Adreno
>
> Original change's description:
> > Simplify _blend_set_color_saturation, removing an instruction.
> >
> > This tightens up our intrinsics slightly; after inlining, it eliminates
> > one scratch variable. (We no longer need to copy `sda` into `hueColor`
> > as hueColor is now unchanged.)
> >
> > Change-Id: Iece5ba2fe11cde54481704a1787114a2c2a66d9b
> > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336599
> > Commit-Queue: John Stiles <johnstiles@google.com>
> > Commit-Queue: Brian Osman <brianosman@google.com>
> > Auto-Submit: John Stiles <johnstiles@google.com>
> > Reviewed-by: Brian Osman <brianosman@google.com>
>
> TBR=brianosman@google.com,johnstiles@google.com
>
> Change-Id: Ica506467b0a4e03d0cbe482034acfa2d9f8d2c16
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337560
> Reviewed-by: John Stiles <johnstiles@google.com>
TBR=brianosman@google.com,johnstiles@google.com
Change-Id: Ia93263f3269c057e7eaa69ca2b05e783d18c0199
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337944
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, we'd gauge suitability for inlining by counting the nodes in
a function; past a certain limit, the function was considered "too big."
Now, we also incorporate the number of times that function is called.
So if a function is called three times, and its size is 20 nodes, it
would be considered to have an inlining cost of 60 (3 * 20) instead of
20.
This should tamp down the aggressive nature of the inliner in cases like
gaussian convolution or complicated blends, and will hopefully satisfy
Pinpoint.
No change visible in Nanobench (which doesn't test any of these sorts of
patterns, but certainly inlines things): http://screen/AwD5hkgkEfjVx4g
Change-Id: Ie5e32898245ac854adb9ddd52d87001df6a67125
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337676
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
- float4(float2(1, 2), 3, 4) --> float4(1, 2, 3, 4)
- half3(z, half2(fn(x), y*2)) --> half3(z, fn(x), y*2)
Single-argument constructors will be ignored by this optimization; these
might be casts or splats.
This had an unexpected side benefit of simplifying some Metal output,
as we need to output fewer Metal matrix construction helper functions
when matrices use more simple scalars for construction.
Change-Id: I0a161db060c107e35247901619291bf83801cb11
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337400
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, every output was labeled ".asm.frag" regardless of the
actual type.
Change-Id: Icf3a56bb04d88cc0443f12c2dfb99c66ee00dff0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337717
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Change-Id: I1be21b428939d17bbf3a9347a64db56c7cd69eb4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337638
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This reverts commit ed289e777c.
Reason for revert: causing strange artifacts, only on Adreno
Original change's description:
> Simplify _blend_set_color_saturation, removing an instruction.
>
> This tightens up our intrinsics slightly; after inlining, it eliminates
> one scratch variable. (We no longer need to copy `sda` into `hueColor`
> as hueColor is now unchanged.)
>
> Change-Id: Iece5ba2fe11cde54481704a1787114a2c2a66d9b
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336599
> Commit-Queue: John Stiles <johnstiles@google.com>
> Commit-Queue: Brian Osman <brianosman@google.com>
> Auto-Submit: John Stiles <johnstiles@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
TBR=brianosman@google.com,johnstiles@google.com
Change-Id: Ica506467b0a4e03d0cbe482034acfa2d9f8d2c16
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337560
Reviewed-by: John Stiles <johnstiles@google.com>
These are not actually supported operators in GLSL, Metal or SPIR-V and
we don't emulate them. Their absence was causing SPIR-V to fail the
Operators.sksl test.
Change-Id: Ia6933788392aea48836b7be19e32b9969805f254
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337185
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, the code which calculated Constructor constant values
assumed that a constant-value PrefixExpression would always have an
operand of Constructor. It turns out that another valid case is multiple
PrefixExpressions nested within each other (representing repeated
negation). Updated the code to work regardless of the type of the prefix
operand.
Change-Id: Ic9bf54725ae59330ac817bc4ec7a64def384ab54
Bug: oss-fuzz:27663
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337177
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
We now have SPIR-V golden outputs for `blend` and `shared` tests.
This exposes a handful of SPIR-V limitations for us to address.
Change-Id: Ie5278889b8a61432403d06231b17765885bee0ac
Bug: skia:10694
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337182
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This allows swizzle removal to apply in more cases; in particular, we
can now optimize away extra swizzles caused by zero/one swizzle-
components quite effectively.
The "trivial expression" code was lifted from the inliner. Some subtle
changes in trivial-expression determination affect the inliner's results
in boring, non-meaningful ways. In particular, multi-argument
constructors containing all-constant values are now considered trivial,
whereas previously only single-argument constructors made the trivial-
ness cut. This allows the inliner to propagate some values that it
wouldn't have before.
Change-Id: I9a009b6803d9ac9595d65538252ba81c2b7166a7
Bug: skia:10954
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336156
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, any builtin functions would be optimized as a side-effect of
optimizing programs that used them. Now that shared elements aren't
being optimized in that way, we explicitly optimize any shared modules
when they are first created. We don't remove dead elements, but we
we do substitute settings, simplify, and inline.
Bug: skia:10905
Change-Id: I701b5e9f52fb880ef3e6f4c67694d08602f47e95
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336440
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
The fix submitted at http://review.skia.org/335868 did not support
casts. The fuzzer discovered this shortcoming right away.
Change-Id: I2f5166528cee41367348564d4e664476fd5704ff
Bug: oss-fuzz:27650
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336656
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This tightens up our intrinsics slightly; after inlining, it eliminates
one scratch variable. (We no longer need to copy `sda` into `hueColor`
as hueColor is now unchanged.)
Change-Id: Iece5ba2fe11cde54481704a1787114a2c2a66d9b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336599
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
When values from the same argument are used consecutively by the outer
swizzle, they can be merged in the inner swizzle. Merging isn't always
possible, of course, but it will be used where it can be:
`half4(1, colRGB).yzwx` --> `half4(colRGB.xyz, 1)`
`half4(1, colRGB).yxzw` --> `half4(colRGB.x, 1, colRGB.yz)`
Change-Id: Id164b046bc15022ded331c06d722f1ae3605a3bd
Bug: skia:10954
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335872
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This will reorder constructors with swizzles applied, such as
`half4(1, 2, 3, 4).xxyz` --> `half4(1, 1, 2, 3)`
`half4(1, colRGB).yzwx` --> `half4(colRGB.x, colRGB.y, colRGB.z, 1)`
Note that, depending on the swizzle components, some elements of the
constructor may be duplicated and others may be eliminated. The
optimizer makes sure to leave the swizzle alone if it would duplicate
anything non-trivial, or if it would eliminate anything with a side
effect.
Change-Id: I470fda217ae8cf5828406b89a5696ca6aebf608d
Bug: skia:10954
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335860
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This was found at https://oss-fuzz.com/testcase-detail/5155684475469824
but the associated oss-fuzz issue ID appears to be misdirected (it's
showing oss-fuzz:24498, an unrelated issue).
PrefixExpressions can return true for `isCompileTimeConstant` but did
not implement `compareConstant`; the fuzzer discovered this. Because
compile-time constants can only be compared if they are of the same
kind, this means that `compareConstant` is actually comparing a pair of
expressions that are both negated. These negations will just cancel
out, so `compareConstant` on a pair of PrefixExpressions can just call
`compareConstant` on the inner operand of each expression.
Change-Id: I7793e25314e6c8a74278b73299d310794baf71f4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335870
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The fuzzer managed to create a test case which temporarily evaluates to
expression `half2(half(0.2)) + 2` as it is optimized. This requires a
bunch of temporary nonsense math as the IR Generator is attempting to
simplify as it goes; various attempts to remove terms from the fuzzer
test-case would cause it to stop reproducing the error.
Constructor::getVecComponent assumed that any constructor with a single
scalar argument would always implement `getConstantFloat` and
`getConstantInt`; however, constructors themselves did not actually
implement these methods. This meant that nesting a scalar constructor
inside a non-scalar constructor would abort when it tried to deduce the
value inside the inner constructor.
This has been fixed by implementing `getConstantFloat` and
`getConstantInt` for Constructors. These methods will assert if the
constructor has more than one argument or is a non-scalar type. This
should allow any number of nested constructors, e.g.
`half4(half(half(half(1))))` should recursively evaluate properly,
should we somehow generate this as an intermediate expression.
Change-Id: Iaee4284cba03974443cd7b5dccfd7909c1a5f3a6
Bug: oss-fuzz:27614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335868
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The optimizer can now turn the expression `half4(1).xyz` into
`half3(1)`, or `half4(1).w` into `1`. This is actually a somewhat common
case when inlining chains of fragment processors, as inputs are often
overridden to `half4(1)` or `half4(0)`. This optimization also applies
to more complex cases, e.g.:
`half2(anyFunc(sqrt(2))).yxyx` --> `half4(anyFunc(sqrt(2)))`
Since the interior of the constructor is always evaluated once in either
case, it does not actually matter what the constructor contains.
Change-Id: I8d5f358502eaa8e35d4968e74fbd6b0ce2ab6365
Bug: skia:10954
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335818
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This was slightly complicated by the fact that this syntax indicates an
array with a known size:
float[] x = float[](1, 2, 3, 4);
Of course, the size is 4; it's just never explicitly stated in the
code. (The SkSL parser never actually deduces the size, but it doesn't
apparently have a need to; we don't do much in the way of optimization
for arrays.) However, this prevents us from simply failing whenever we
parse "[]" in non-builtin code; we need to keep scanning and see if the
variable is initialized. We already check this in the
ArrayConstructors.sksl test file.
Change-Id: I5b86958e81bd9bf5edf28a617cecf95c1875583e
Bug: skia:10957
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335240
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, the Type's fOffset was set to -1 during parsing, so any
errors related to the Type would be reported on line 1.
Change-Id: I9834f733bc763c5946b3ff81d8aef4807cdc13d1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335584
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This is a followup to http://review.skia.org/335196. This detects opaque
types (samplers and textures) at parsing or IR generation time and
reports an error regardless of backend. This check occurs before Metal
or SPIR-V would have a chance to detect the error, so it changes their
output to a slightly more focused error message. The Metal/SPIR-V fix in
the prior CL is still a nice broad catch-all for preventing spurious
ABORTs, though.
Change-Id: I4cce92a8767d72b5d3d7277a8afde8ce5ce86db2
Bug: skia:10956
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/335217
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
