I ran into a snag while trying to hook up SkRuntimeEffect with debug
tracing. Runtime effects only have access to a skvm::Builder, and are
never exposed to the full skvm::Program. SkVMBlitter is responsible for
assembling the full skvm::Program, but is oblivious to runtime effects
and nested skvm sub-programs. Additionally, multiple runtime effects can
(and often do) coexist within a paint.
This CL changes how debug traces are enabled. skvm::Program no longer
has a `attachDebugTrace` method. Instead, this method lives on the
skvm::Builder. Calling `attachDebugTrace` generates a "trace-hook ID"
(which is actually an index into a vector of TraceHook pointers).
Every trace opcode now includes this trace hook ID. When the Builder
assembles a final Program, it copies the TraceHooks into the Program.
The skvm interpreter uses the trace hook ID on the op to dispatch a
trace command to its associated TraceHook.
From a user perspective, this doesn't change very much, but it does
mean that the SkVM Code Generator now supplies a TraceHook for us
(since it adds trace ops to the Builder and needs to know the proper
trace-hook ID).
Change-Id: I8bd5fea24f477f81470fae8ba41be45f76949407
Bug: skia:12708
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/479597
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, the trace opcodes took a single mask argument, which was
computed as `execution mask & trace mask`. This led to extra bit_ands in
the output, as this value would be need to be recalculated every time
the execution mask changed.
To reduce this cost on program size, the trace ops now take two mask
arguments and require that both must be true. We have four register
slots at our disposal in an Op, which is more than we need, so this
doesn't really cost us anything.
(As an extra minor optimization, if one of the masks is "always-on", we
optimize it away. This avoids burning a register just to hold a ~0
immediate value.)
Change-Id: I9eb71292a1983e71b03c7ac842534beb3d6bbf17
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/478456
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
switch is no longer an ES3-specific feature.
Change-Id: Ic878a77268e517e17699c2e35a37da6b0a7765dd
Bug: skia:12450
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/452320
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This demonstrates how much additional non-trace code can be generated
when debug traces are turned on. A followup CL adds a setting to disable
`trace_var`, which eliminates a significant percentage of the additional
ops (at the cost of removing valuable debug info).
Change-Id: I238e28e6f6531f1dbccfef8f1dcd24a1e8481669
Bug: skia:12692
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/478416
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Julia Lavrova <jlavrova@google.com>
This shouldn't change code generation when debug traces are disabled.
When they are enabled, we now get trace_var opcodes emitted for every
return statement. Internally, this required a fair amount of refactoring
around how return values are passed around, but it should all be
functionally equivalent.
Change-Id: Ieb9d9c75399109186f905e0499d8fe6e2fc2067c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/477981
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
SkRuntimeEffect users will be insulated from skvm and the Builder, so
they won't be able to make an skvm::Coord directly. Replace this with
an SkIPoint, and do the conversion inside the SkVM code generator.
This also removes the requirement of assigning a trace coordinate; if
no trace coordinate is set, (0, 0) will be used.
Change-Id: I97be025d0ea146a86655cc3cdb76726f3bbc2324
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/474556
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reusing the same opcode for enter and exit didn't have any real upside,
and forced us to deal with a fake immediate-value.
Change-Id: I35cab4d196f55b6f5e3040074c731b298720d62b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/477157
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
I should have realized the fuzzer would find this assert when I added
it. Now the front-end rejects these layout qualifiers on both struct
fields and interface block fields. LayoutInInterfaceBlock.sksl is a
reformatted version of the fuzzer input. LayoutInStruct is hand-crafted
to trigger the same failure on a different code path. Both would
previously assert in the SPIRV generator. Now, neither one gets that
far.
Bug: oss-fuzz:41347
Change-Id: Iff69d8f5482da7b772e9331c4fd2d58e89813c46
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/476396
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This reverts commit b1f450bb2b.
Reason for revert: Hmm, bot disagrees.
Original change's description:
> Update some skvm test outputs
>
> Not sure how these were wrong in the repo for so long - these all should
> have changed like this when the optimization of commutative operations
> landed: reviews.skia.org/473239
>
> Change-Id: I3035d4f29c208096f83184b4393023666e10ab92
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/476316
> Auto-Submit: Brian Osman <brianosman@google.com>
> Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Change-Id: I9dc808927c73a979f4cde5c7035b109d8db808f6
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/476397
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Not sure how these were wrong in the repo for so long - these all should
have changed like this when the optimization of commutative operations
landed: reviews.skia.org/473239
Change-Id: I3035d4f29c208096f83184b4393023666e10ab92
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/476316
Auto-Submit: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This is a reland of 9372ef0228
Original change's description:
> Restrict where 'binding' and 'set' can appear
>
> In SPIRV, these are an error when applied to struct members. Some of our
> tests were triggering that because we had free-floating uniforms
> decorated this way (and we coalesce those into members of an interface
> block).
>
> Now, we only allow those layout qualifiers on variable types that will
> remain top-level constructs in the back-end.
>
> Bug: skia:12670
> Change-Id: I73e69cecf6237a1c1180ad38d9b5d52ea80316fb
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/474218
> Commit-Queue: Brian Osman <brianosman@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Bug: skia:12670
Change-Id: I01c0323bba7ce0bddea5f9fb907e2b60e6b812d2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/475156
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This reverts commit 9372ef0228.
Reason for revert: Unhappy bots
Original change's description:
> Restrict where 'binding' and 'set' can appear
>
> In SPIRV, these are an error when applied to struct members. Some of our
> tests were triggering that because we had free-floating uniforms
> decorated this way (and we coalesce those into members of an interface
> block).
>
> Now, we only allow those layout qualifiers on variable types that will
> remain top-level constructs in the back-end.
>
> Bug: skia:12670
> Change-Id: I73e69cecf6237a1c1180ad38d9b5d52ea80316fb
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/474218
> Commit-Queue: Brian Osman <brianosman@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Bug: skia:12670
Change-Id: Ie518192d9a52fc896e615ec08ce0674ad683ec61
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/475099
Auto-Submit: Brian Osman <brianosman@google.com>
Commit-Queue: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
In SPIRV, these are an error when applied to struct members. Some of our
tests were triggering that because we had free-floating uniforms
decorated this way (and we coalesce those into members of an interface
block).
Now, we only allow those layout qualifiers on variable types that will
remain top-level constructs in the back-end.
Bug: skia:12670
Change-Id: I73e69cecf6237a1c1180ad38d9b5d52ea80316fb
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/474218
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This is redundant information now that we have SkVMSlotInfo.
Change-Id: Ia05b1eaa722023e719042c83255708aa9debed61
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/473777
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This adds new validation rules that we were breaking.
Binding and DescriptorSet can't be applied to push constants, nor to
struct members.
Bug: skia:12670
Bug: chromium:1270328
Change-Id: I332f77717b08d9945c8e5b79c5bf649a8f5f2043
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/474056
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Previously, we would avoid emitting redundant trace_vars by checking to
see if the slot is being assigned to the exact same Var.ID. However,
this had the potential to eliminate useful trace_vars:
- At the start of execution, slots all contain 0. Code which explicitly
assigned a zero into a slot would not be shown in a trace. (So things
like `color = half4(0,1,0,1)` would only emit traces for color.ga.)
- A function call's parameter slots are reused every time it is called,
so calling a function twice would only emit traces for the parameters
that aren't the same Val.ID as the previous call.
- A VarDeclaration inside a loop reuses its slot each time through the
loop, even though conceptually it's a "new" variable.
We now track a slot's "written-to" status. At the start of execution,
no slots have been "written-to". These slots will always emit a
trace_var opcode (fixing the first issue). Also, issuing a function
call or declaring a variable will reset the "written-to" status of the
associated slots (fixing the second and third issues).
When the debugger is not in use, the written-to field is unused.
Change-Id: I482a86cb6e90d0f85dd2a161e984f212782a7b4d
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/473776
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
We now canonicalize commutative operations by ordering their value IDs.
The lower-numbered value ID is always placed first into a commutative
instruction. In other words, this instruction:
bit_and result, v7, v5
Would be silently converted to this:
bit_and result, v5, v7
This will allow these two logically-equivalent instructions to be
deduplicated:
bit_and result, v7, v5
bit_and result, v5, v7
Of course, deduplicating these ops can unlock additional free CSE/DCE.
The affected instructions are listed in http://review.skia.org/473238
Change-Id: Ib9beb79d6b72d7903184aaa9a53e8e5a02ae126d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/473239
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Herb Derby <herb@google.com>
The SkVMDebugInfo now includes an skvm::Coord that the caller is
responsible for filling in before converting the program. If it is set,
the value indicates the device coordinates that should be traced. If it
is unset, debug traces will not be emitted at all.
Within the SkVMCodeGenerator, we now have a new traceMask() call which
combines the current execution mask with the trace mask. Tracing opcodes
now pass the result of traceMask() instead of mask(). This will limit
trace data to the selected pixel, instead of tracing the entire draw.
Bug: skia:12614
Change-Id: I1f8ce7d18a31aa60a139a4a7335c2a285d6aee60
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472798
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
We no longer indicate functions by their line number, which can be
ambiguous. The debug info now includes a list of function names which we
can refer to by index, and the `trace_call` opcode references functions
by their index in this list.
Change-Id: I4bdf2a6439637a07b116831fd2981f342c19ea4a
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472796
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This fails exactly as it should, but we had no test for it.
Change-Id: I0aa3307c444f2c9bc3512ff43b784a56a7c09856
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472449
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This wasn't meaningful, and made some error reporting worse.
Change-Id: I5e72b5aca5d3e159b8439fa9809290d75e44cbe2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472656
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
The `eval` methods take a shader/blender/colorFilter, and we assumed
when assembling the ChildCall expression that the child expression would
be a VariableReference because opaque objects don't participate in
normal expressions. However, comma-expressions were allowed to contain
opaque types. GLSL doesn't allow opaque types in comma-expressions:
http://screen/8YW59tYDUbBh9eW
Now we disallow them as well.
Change-Id: Iaf88ef7bddb5cc8f1f1e23b515174dfc291e00c7
Bug: oss-fuzz:41072
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472446
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
`void` behaves differently from opaque types in several situations, such
as function return types, and errors involving `void` should not call it
an opaque type. I've fixed all the places where we relied on `isOpaque`
to return catch void types.
Change-Id: I359c00f836be6e56cb0373beff60cf2ff830f77b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472451
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Mysteriously, I had written a test which put arrays of void inside a
struct, but had neglected to include the non-array case. It causes an
okay-not-great error (referring to void as an "opaque type").
Change-Id: Id20a9d3512d29aecea81d46877dce708b7b2f973
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472450
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Change-Id: I209119e6c74ca54dd6021b6dec4775fc7b66adeb
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472448
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
We should, of course, detect this and report an error.
Change-Id: I42b3be6e714a1f367d3251842506a384f2afe019
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472447
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, we'd report them as an overflowed integer.
Change-Id: Ia3632b4bc880829fb04b08a002d7ce9523567a54
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/472056
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This is required by the ES2 standard: http://screen/Qysv4fPW5r5LA9e
This actually already worked fine because `strtoull` natively recognizes
octal values without any work on our part. However, we lacked a test.
Change-Id: I3033de899918abe99c63a9b7b79bd4c3374ee315
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/471716
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The only type of expressions that getConstantSubexpression could ever
return are Literal and nullptr. getConstantValue now returns an
optional<double>; nullopt indicates a non-constant value in the slot.
This simplifies most use cases, and allows us to get rid of some extra
"zero" and "one" Literal objects in some of our Constructor classes.
This change fixes a recent fuzzer issue. The fuzzer had discovered that
calling `getConstantSubexpression` on a ConstructorCompoundCast that
contained a compile-time-constant value would return literals of the
wrong type (the cast was not applied). By nesting repeated matrix casts,
this type confusion could be turned into an assertion.
Change-Id: Icee69219e6db2822ffdfab4e5ccdaff54584a4b6
Bug: oss-fuzz:41000
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/471376
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This assigns a human-readable name to a debug slot. The slot map is
emitted into skslc output files, and will be used in the future to
display human-readable names in the debugger.
Change-Id: I288358de305239005faa5814bd1d77a38b5e05b0
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470400
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 3aaed99930.
Reason for revert: removing changes to PrecisionQualifiers
Original change's description:
> Revert "Fix Metal codegen error with structs containing compound types."
>
> This reverts commit 2a6c41571b.
>
> Reason for revert: causing Mali G7x failures on tree
>
> Original change's description:
> > Fix Metal codegen error with structs containing compound types.
> >
> > While working on an unrelated test, I accidentally triggered a bug in
> > Metal code generation. Our struct-equality helper functions did not
> > properly handle vector fields. Wrapping each comparison in `all(...)`
> > fixes the problem. (all() on a scalar is allowed and does nothing.)
> >
> > Our struct comparison tests now include a vector and a matrix.
> >
> > Change-Id: I59061ae9c3c3ab2c2dbdcb5257bc23e2257152af
> > Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470399
> > Commit-Queue: John Stiles <johnstiles@google.com>
> > Auto-Submit: John Stiles <johnstiles@google.com>
> > Reviewed-by: Brian Osman <brianosman@google.com>
>
> TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com,skcq-be@skia-corp.google.com.iam.gserviceaccount.com
>
> Change-Id: Ieb5d5a1839978fb82525863488e9d54fdf44adbd
> No-Presubmit: true
> No-Tree-Checks: true
> No-Try: true
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/471097
> Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
> Reviewed-by: Chris Dalton <csmartdalton@google.com>
> Commit-Queue: John Stiles <johnstiles@google.com>
Change-Id: I8ee90df3de075cf82c0fcf3b4787577b09bb1a70
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/471156
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This reverts commit 2a6c41571b.
Reason for revert: causing Mali G7x failures on tree
Original change's description:
> Fix Metal codegen error with structs containing compound types.
>
> While working on an unrelated test, I accidentally triggered a bug in
> Metal code generation. Our struct-equality helper functions did not
> properly handle vector fields. Wrapping each comparison in `all(...)`
> fixes the problem. (all() on a scalar is allowed and does nothing.)
>
> Our struct comparison tests now include a vector and a matrix.
>
> Change-Id: I59061ae9c3c3ab2c2dbdcb5257bc23e2257152af
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470399
> Commit-Queue: John Stiles <johnstiles@google.com>
> Auto-Submit: John Stiles <johnstiles@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com,skcq-be@skia-corp.google.com.iam.gserviceaccount.com
Change-Id: Ieb5d5a1839978fb82525863488e9d54fdf44adbd
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/471097
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
While working on an unrelated test, I accidentally triggered a bug in
Metal code generation. Our struct-equality helper functions did not
properly handle vector fields. Wrapping each comparison in `all(...)`
fixes the problem. (all() on a scalar is allowed and does nothing.)
Our struct comparison tests now include a vector and a matrix.
Change-Id: I59061ae9c3c3ab2c2dbdcb5257bc23e2257152af
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470399
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, none of our `runtime` tests relied on the input coordinate
in any way, so all of the logic was hoisted above the main loop in every
test. This CL adds an artificial reliance on the input coordinate so
that we have at least some SkVM tests with real code in the main loop.
This lets us see debug trace instructions interleaved with real code.
The input coordinate is clamped against a known uniform value
(`colorGreen` always contains 0101) so that the final test output
remains consistent in practice.
Additionally, I noticed that this test was only enabled in ES3, but
it doesn't seem to have anything ES3-specific in it, so it's now
enabled across the board.
Change-Id: Ie82f40b1060edb6071e300040ac59fb7d27094b0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470397
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Our inliner would ignore any functions with `inout` parameters, because
inlining them properly was more complex than just leaving the function
call. However, real-world code can sometimes contain helper functions
that have `inout` params that are never used at all (e.g. an uber-shader
with some features turned off).
We now read the ProgramUsage and check to see whether or not the
`inout`-qualified parameter is actually modified. If it's never changed,
the function now remains a candidate for inlining.
Change-Id: I92e494f94cc070801cb9aa28bd13faa689b806b6
Bug: skia:12636
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470299
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Yesterday's implementation was close but I realized later that it wasn't
quite ideal.
- Array index-folding was gated on `isCompileTimeConstant`, which is too
strict. The real limitation is `hasSideEffects`. If an array contains
a side-effecting expression, we should leave it alone. Otherwise it
is safe to pluck out an element from the array and toss the rest.
- Matrix index-folding was gated on `getConstantSubexpression` for the
extracted elements, but did not check the other elements at all. This
was too lenient; we now only proceed to the folding step if
`hasSideEffects` returns false.
I added some tests to verify the final behavior and also discovered a
small related issue. Diagonal matrices were not substituting literals
in for constant-values, which inhibited folding as well and would break
constant-expression evaluation. This is now fixed.
Change-Id: Idda32fd8643c1f32ba21475251cd4d4dd7cea94c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470396
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Functions which don't write to their out params should be safe to
inline, but we currently don't recognize this.
Change-Id: I753e48067c7be4473675ef6c95e61af17dc5ae41
Bug: skia:12636
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470298
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
We never used it internally, but the shaders used by Filament rely on
it. It doesn't exist in ES2 so this doesn't affect Runtime Effects.
Change-Id: Idb2afb15ff160b950ad02101bf6381a5d5c56468
Bug: skia:12635, skia:11209
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/470156
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Indexing into a constant matrix is a constant expression, so we are
obligated to support it for ES2 compatibility.
Change-Id: Ibe1e5bac39d9a88ce0222997a38e8b6952fdb336
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469819
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
We should support constant-expressions involving matrices (GLSL ES2
does, WebGL does). We currently don't. We do properly report out-of-
range indexing, but we don't optimize away valid matrix index
expressions or allow matrices to be indexed in a constant-expression
context.
Change-Id: If58aa4c5f15abef421a412957072f3617b4176df
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469818
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, SkSL was unable to resolve the constant expression `x[y]`
for a constant-array `x` and a constant-integer-scalar `y`. Now, if `x`
and `y` are known, we can replace `x[y]` with the indexed array element.
Note that we need to be careful here, as it's not a valid optimization
to eliminate array elements that have side effects. We preserve side-
effecting expressions using the comma operator.
Change-Id: I5721337eb42b48c0b05f919c1cadfae19dd3b84f
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469839
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, we didn't have tests which leveraged constant-evaluation of
array indexing (because we didn't support it), and our test files
commingled constant-indexing into vectors with constant-indexing into
arrays.
The test files now separate vector- and array-handling into separate
tests, and a ton of new cases have been added to ArrayFolding. The
ArrayFolding tests now require constant-evaluation of array indexing,
so they fail in this CL, but will be fixed in the followup CL.
Change-Id: I3b663e743d97d6db80627bc9b7808f88c99917a7
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469528
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, this code assumed that IndexExpression::Convert had done
range checking and that it was safe to access the base expression at
the passed-in index. The inliner violates this assumption, because it
can replace unknowns (where out-of-range access is undefined but non-
fatal) with knowns (where out-of-range access is forbidden).
We now do range-checking inside IndexExpression::Make and report the
error cleanly, instead of asserting inside of Swizzle::Make due to an
invalid component index.
Change-Id: If0f31b1f694bcc2a875d124f70be311d6634c77b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469535
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, a dangling type or function reference would be eliminated
silently with optimizations on, or would assert when optimizations were
off.
Change-Id: Ib2e273b6f069724e8872c9cb97351b647b875a62
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469525
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The ExpressionStatement currently eliminates dangling references without
reporting them as an error. This happens due to optimization; these
expressions (being meaningless) have no side effects, and so the
optimizer replaces them with Nop. When the optimizer is off, these
programs trigger an assert:
https://osscs.corp.google.com/skia/skia/+/main:src/sksl/SkSLAnalysis.cpp;l=582;drc=e7a953524787e3bd0c437ec52de4e40986689825
A followup CL will fix ExpressionStatements so that they report
incomplete expressions as an error.
Change-Id: Ica49166032e670749fc1b4e7a869fbab03364d4f
Bug: skia:12472
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469524
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This enables stepping over function calls automatically.
Change-Id: Ie15ed745377d851cb7752f651b573efa2cc8195f
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469077
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Previously, the for statement's "increment/test" expressions were
executed without moving the trace-line back up to the for statement.
When stepping through code, we will now explicitly step to the next/test
line on each loop iteration.
Change-Id: I5d9f005a42150670cec77218323cf932ee1cbdb0
Bug: skia:12614
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/469180
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
