Google Git
Sign in
skia / skia / f30529658275b40e98fe140aa20439057b49f8d7 / . / src / sksl / codegen / SkSLRasterPipelineCodeGenerator.cpp
blob: c1d3533745526116108fe4350fd8587138c9d7af [file] [log] [blame]
/*
* Copyright 2022 Google LLC
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#include "include/core/SkSpan.h"
#include "include/private/SkSLDefines.h"
#include "include/private/SkSLIRNode.h"
#include "include/private/SkSLLayout.h"
#include "include/private/SkSLModifiers.h"
#include "include/private/SkSLProgramElement.h"
#include "include/private/SkSLStatement.h"
#include "include/private/SkSLString.h"
#include "include/private/base/SkTArray.h"
#include "include/sksl/SkSLOperator.h"
#include "include/sksl/SkSLPosition.h"
#include "src/base/SkStringView.h"
#include "src/core/SkTHash.h"
#include "src/sksl/SkSLAnalysis.h"
#include "src/sksl/SkSLBuiltinTypes.h"
#include "src/sksl/SkSLCompiler.h"
#include "src/sksl/SkSLIntrinsicList.h"
#include "src/sksl/codegen/SkSLRasterPipelineBuilder.h"
#include "src/sksl/codegen/SkSLRasterPipelineCodeGenerator.h"
#include "src/sksl/ir/SkSLBinaryExpression.h"
#include "src/sksl/ir/SkSLBlock.h"
#include "src/sksl/ir/SkSLBreakStatement.h"
#include "src/sksl/ir/SkSLChildCall.h"
#include "src/sksl/ir/SkSLConstructor.h"
#include "src/sksl/ir/SkSLConstructorCompound.h"
#include "src/sksl/ir/SkSLConstructorDiagonalMatrix.h"
#include "src/sksl/ir/SkSLConstructorMatrixResize.h"
#include "src/sksl/ir/SkSLConstructorSplat.h"
#include "src/sksl/ir/SkSLContinueStatement.h"
#include "src/sksl/ir/SkSLDoStatement.h"
#include "src/sksl/ir/SkSLExpression.h"
#include "src/sksl/ir/SkSLExpressionStatement.h"
#include "src/sksl/ir/SkSLFieldAccess.h"
#include "src/sksl/ir/SkSLForStatement.h"
#include "src/sksl/ir/SkSLFunctionCall.h"
#include "src/sksl/ir/SkSLFunctionDeclaration.h"
#include "src/sksl/ir/SkSLFunctionDefinition.h"
#include "src/sksl/ir/SkSLIfStatement.h"
#include "src/sksl/ir/SkSLIndexExpression.h"
#include "src/sksl/ir/SkSLLiteral.h"
#include "src/sksl/ir/SkSLPostfixExpression.h"
#include "src/sksl/ir/SkSLPrefixExpression.h"
#include "src/sksl/ir/SkSLProgram.h"
#include "src/sksl/ir/SkSLReturnStatement.h"
#include "src/sksl/ir/SkSLSwizzle.h"
#include "src/sksl/ir/SkSLTernaryExpression.h"
#include "src/sksl/ir/SkSLType.h"
#include "src/sksl/ir/SkSLVarDeclarations.h"
#include "src/sksl/ir/SkSLVariable.h"
#include "src/sksl/ir/SkSLVariableReference.h"
#include "src/sksl/tracing/SkRPDebugTrace.h"
#include "src/sksl/tracing/SkSLDebugInfo.h"
#include <cstddef>
#include <cstdint>
#include <float.h>
#include <numeric>
#include <optional>
#include <string>
#include <string_view>
#include <utility>
#include <vector>
namespace SkSL {
namespace RP {
class LValue;
class SlotManager {
public:
SlotManager(std::vector<SlotDebugInfo>* i) : fSlotDebugInfo(i) {}
/** Used by `create` to add this variable to SlotDebugInfo inside SkRPDebugTrace. */
void addSlotDebugInfoForGroup(const std::string& varName,
const Type& type,
Position pos,
int* groupIndex,
bool isFunctionReturnValue);
void addSlotDebugInfo(const std::string& varName,
const Type& type,
Position pos,
bool isFunctionReturnValue);
/** Creates slots associated with an SkSL variable or return value. */
SlotRange createSlots(std::string name,
const Type& type,
Position pos,
bool isFunctionReturnValue);
/**
* Creates a single temporary slot for scratch storage. Temporary slots can be recycled, which
* frees them up for reuse. Temporary slots are not assigned a name and have an arbitrary type.
*/
SlotRange createTemporarySlot(const Type& type);
void recycleTemporarySlot(SlotRange temporarySlot);
/** Looks up the slots associated with an SkSL variable; creates the slot if necessary. */
SlotRange getVariableSlots(const Variable& v);
/**
* Looks up the slots associated with an SkSL function's return value; creates the range if
* necessary. Note that recursion is never supported, so we don't need to maintain return values
* in a stack; we can just statically allocate one slot per function call-site.
*/
SlotRange getFunctionSlots(const IRNode& callSite, const FunctionDeclaration& f);
/** Returns the total number of slots consumed. */
int slotCount() const { return fSlotCount; }
private:
std::string makeTempName() { return SkSL::String::printf("[temporary %d]", fTemporaryCount++); }
SkTHashMap<const IRNode*, SlotRange> fSlotMap;
SkTArray<Slot> fRecycledSlots;
int fSlotCount = 0;
int fTemporaryCount = 0;
std::vector<SlotDebugInfo>* fSlotDebugInfo;
};
class AutoContinueMask {
public:
AutoContinueMask() = default;
~AutoContinueMask() {
if (fSlotRange) {
fSlotManager->recycleTemporarySlot(*fSlotRange);
*fSlotRange = fPreviousSlotRange;
}
}
void enable(SlotManager* mgr, const Type& type, SlotRange* range) {
fSlotManager = mgr;
fSlotRange = range;
fPreviousSlotRange = *fSlotRange;
*fSlotRange = fSlotManager->createTemporarySlot(type);
}
void enterLoopBody(Builder& builder) {
if (fSlotRange) {
builder.zero_slots_unmasked(*fSlotRange);
}
}
void exitLoopBody(Builder& builder) {
if (fSlotRange) {
builder.reenable_loop_mask(*fSlotRange);
}
}
private:
SlotManager* fSlotManager = nullptr;
SlotRange* fSlotRange = nullptr;
SlotRange fPreviousSlotRange;
};
class Generator {
public:
Generator(const SkSL::Program& program, SkRPDebugTrace* debugTrace)
: fProgram(program)
, fDebugTrace(debugTrace)
, fProgramSlots(debugTrace ? &debugTrace->fSlotInfo : nullptr)
, fUniformSlots(debugTrace ? &debugTrace->fUniformInfo : nullptr) {}
/** Converts the SkSL main() function into a set of Instructions. */
bool writeProgram(const FunctionDefinition& function);
/** Returns the generated program. */
std::unique_ptr<RP::Program> finish();
/**
* Converts an SkSL function into a set of Instructions. Returns nullopt if the function
* contained unsupported statements or expressions.
*/
std::optional<SlotRange> writeFunction(const IRNode& callSite,
const FunctionDefinition& function);
/**
* Returns the slot index of this function inside the FunctionDebugInfo array in SkRPDebugTrace.
* The FunctionDebugInfo slot will be created if it doesn't already exist.
*/
int getFunctionDebugInfo(const FunctionDeclaration& decl);
/** Looks up the slots associated with an SkSL variable; creates the slot if necessary. */
SlotRange getVariableSlots(const Variable& v) {
SkASSERT(!IsUniform(v));
return fProgramSlots.getVariableSlots(v);
}
/** Looks up the slots associated with an SkSL uniform; creates the slot if necessary. */
SlotRange getUniformSlots(const Variable& v) {
SkASSERT(IsUniform(v));
return fUniformSlots.getVariableSlots(v);
}
/**
* Looks up the slots associated with an SkSL function's return value; creates the range if
* necessary. Note that recursion is never supported, so we don't need to maintain return values
* in a stack; we can just statically allocate one slot per function call-site.
*/
SlotRange getFunctionSlots(const IRNode& callSite, const FunctionDeclaration& f) {
return fProgramSlots.getFunctionSlots(callSite, f);
}
/** The Builder stitches our instructions together into Raster Pipeline code. */
Builder* builder() { return &fBuilder; }
/** Appends a statement to the program. */
[[nodiscard]] bool writeStatement(const Statement& s);
[[nodiscard]] bool writeBlock(const Block& b);
[[nodiscard]] bool writeBreakStatement(const BreakStatement& b);
[[nodiscard]] bool writeContinueStatement(const ContinueStatement& b);
[[nodiscard]] bool writeDoStatement(const DoStatement& d);
[[nodiscard]] bool writeExpressionStatement(const ExpressionStatement& e);
[[nodiscard]] bool writeMasklessForStatement(const ForStatement& f);
[[nodiscard]] bool writeForStatement(const ForStatement& f);
[[nodiscard]] bool writeGlobals();
[[nodiscard]] bool writeIfStatement(const IfStatement& i);
[[nodiscard]] bool writeDynamicallyUniformIfStatement(const IfStatement& i);
[[nodiscard]] bool writeReturnStatement(const ReturnStatement& r);
[[nodiscard]] bool writeVarDeclaration(const VarDeclaration& v);
/** Pushes an expression to the value stack. */
[[nodiscard]] bool pushBinaryExpression(const BinaryExpression& e);
[[nodiscard]] bool pushBinaryExpression(const Expression& left,
Operator op,
const Expression& right);
[[nodiscard]] bool pushChildCall(const ChildCall& c);
[[nodiscard]] bool pushConstructorCast(const AnyConstructor& c);
[[nodiscard]] bool pushConstructorCompound(const ConstructorCompound& c);
[[nodiscard]] bool pushConstructorDiagonalMatrix(const ConstructorDiagonalMatrix& c);
[[nodiscard]] bool pushConstructorMatrixResize(const ConstructorMatrixResize& c);
[[nodiscard]] bool pushConstructorSplat(const ConstructorSplat& c);
[[nodiscard]] bool pushExpression(const Expression& e, bool usesResult = true);
[[nodiscard]] bool pushFunctionCall(const FunctionCall& e);
[[nodiscard]] bool pushIndexExpression(const IndexExpression& i);
[[nodiscard]] bool pushIntrinsic(const FunctionCall& c);
[[nodiscard]] bool pushIntrinsic(IntrinsicKind intrinsic, const Expression& arg0);
[[nodiscard]] bool pushIntrinsic(IntrinsicKind intrinsic,
const Expression& arg0,
const Expression& arg1);
[[nodiscard]] bool pushIntrinsic(IntrinsicKind intrinsic,
const Expression& arg0,
const Expression& arg1,
const Expression& arg2);
[[nodiscard]] bool pushLiteral(const Literal& l);
[[nodiscard]] bool pushPostfixExpression(const PostfixExpression& p, bool usesResult);
[[nodiscard]] bool pushPrefixExpression(const PrefixExpression& p);
[[nodiscard]] bool pushPrefixExpression(Operator op, const Expression& expr);
[[nodiscard]] bool pushSwizzle(const Swizzle& s);
[[nodiscard]] bool pushTernaryExpression(const TernaryExpression& t);
[[nodiscard]] bool pushTernaryExpression(const Expression& test,
const Expression& ifTrue,
const Expression& ifFalse);
[[nodiscard]] bool pushDynamicallyUniformTernaryExpression(const Expression& test,
const Expression& ifTrue,
const Expression& ifFalse);
[[nodiscard]] bool pushVariableReference(const VariableReference& v);
/** Pops an expression from the value stack and copies it into slots. */
void popToSlotRange(SlotRange r) { fBuilder.pop_slots(r); }
void popToSlotRangeUnmasked(SlotRange r) { fBuilder.pop_slots_unmasked(r); }
/** Pops an expression from the value stack and discards it. */
void discardExpression(int slots) { fBuilder.discard_stack(slots); }
/** Zeroes out a range of slots. */
void zeroSlotRangeUnmasked(SlotRange r) { fBuilder.zero_slots_unmasked(r); }
/** Expression utilities. */
struct TypedOps {
BuilderOp fFloatOp;
BuilderOp fSignedOp;
BuilderOp fUnsignedOp;
BuilderOp fBooleanOp;
};
static BuilderOp GetTypedOp(const SkSL::Type& type, const TypedOps& ops);
[[nodiscard]] bool unaryOp(const SkSL::Type& type, const TypedOps& ops);
[[nodiscard]] bool binaryOp(const SkSL::Type& type, const TypedOps& ops);
[[nodiscard]] bool ternaryOp(const SkSL::Type& type, const TypedOps& ops);
[[nodiscard]] bool pushIntrinsic(const TypedOps& ops, const Expression& arg0);
[[nodiscard]] bool pushIntrinsic(const TypedOps& ops,
const Expression& arg0,
const Expression& arg1);
[[nodiscard]] bool pushIntrinsic(BuilderOp builderOp, const Expression& arg0);
[[nodiscard]] bool pushIntrinsic(BuilderOp builderOp,
const Expression& arg0,
const Expression& arg1);
[[nodiscard]] bool pushVectorizedExpression(const Expression& expr, const Type& vectorType);
[[nodiscard]] bool pushVariableReferencePartial(const VariableReference& v, SlotRange subset);
[[nodiscard]] bool pushLValueOrExpression(LValue* lvalue, const Expression& expr);
[[nodiscard]] bool pushMatrixMultiply(LValue* lvalue,
const Expression& left,
const Expression& right,
int leftColumns, int leftRows,
int rightColumns, int rightRows);
void foldWithMultiOp(BuilderOp op, int elements);
void nextTempStack() {
fBuilder.set_current_stack(++fCurrentTempStack);
}
void previousTempStack() {
fBuilder.set_current_stack(--fCurrentTempStack);
}
void pushCloneFromNextTempStack(int slots, int offsetFromStackTop = 0) {
fBuilder.push_clone_from_stack(slots, fCurrentTempStack + 1, offsetFromStackTop);
}
void pushCloneFromPreviousTempStack(int slots, int offsetFromStackTop = 0) {
fBuilder.push_clone_from_stack(slots, fCurrentTempStack -1, offsetFromStackTop);
}
BuilderOp getTypedOp(const SkSL::Type& type, const TypedOps& ops) const;
bool needsReturnMask() {
Analysis::ReturnComplexity* complexity = fReturnComplexityMap.find(fCurrentFunction);
if (!complexity) {
complexity = fReturnComplexityMap.set(fCurrentFunction,
Analysis::GetReturnComplexity(*fCurrentFunction));
}
return *complexity >= Analysis::ReturnComplexity::kEarlyReturns;
}
static bool IsUniform(const Variable& var) {
return var.modifiers().fFlags & Modifiers::kUniform_Flag;
}
static bool IsOutParameter(const Variable& var) {
return (var.modifiers().fFlags & (Modifiers::kIn_Flag | Modifiers::kOut_Flag)) ==
Modifiers::kOut_Flag;
}
static bool IsInoutParameter(const Variable& var) {
return (var.modifiers().fFlags & (Modifiers::kIn_Flag | Modifiers::kOut_Flag)) ==
(Modifiers::kIn_Flag | Modifiers::kOut_Flag);
}
private:
const SkSL::Program& fProgram;
Builder fBuilder;
SkRPDebugTrace* fDebugTrace = nullptr;
SkTHashMap<const Variable*, int> fChildEffectMap;
SlotManager fProgramSlots;
SlotManager fUniformSlots;
const FunctionDefinition* fCurrentFunction = nullptr;
SlotRange fCurrentFunctionResult;
SlotRange fCurrentContinueMask;
int fCurrentTempStack = 0;
SkTHashMap<const FunctionDefinition*, Analysis::ReturnComplexity> fReturnComplexityMap;
static constexpr auto kAbsOps = TypedOps{BuilderOp::abs_float,
BuilderOp::abs_int,
BuilderOp::unsupported,
BuilderOp::unsupported};
static constexpr auto kAddOps = TypedOps{BuilderOp::add_n_floats,
BuilderOp::add_n_ints,
BuilderOp::add_n_ints,
BuilderOp::unsupported};
static constexpr auto kSubtractOps = TypedOps{BuilderOp::sub_n_floats,
BuilderOp::sub_n_ints,
BuilderOp::sub_n_ints,
BuilderOp::unsupported};
static constexpr auto kMultiplyOps = TypedOps{BuilderOp::mul_n_floats,
BuilderOp::mul_n_ints,
BuilderOp::mul_n_ints,
BuilderOp::unsupported};
static constexpr auto kDivideOps = TypedOps{BuilderOp::div_n_floats,
BuilderOp::div_n_ints,
BuilderOp::div_n_uints,
BuilderOp::unsupported};
static constexpr auto kLessThanOps = TypedOps{BuilderOp::cmplt_n_floats,
BuilderOp::cmplt_n_ints,
BuilderOp::cmplt_n_uints,
BuilderOp::unsupported};
static constexpr auto kLessThanEqualOps = TypedOps{BuilderOp::cmple_n_floats,
BuilderOp::cmple_n_ints,
BuilderOp::cmple_n_uints,
BuilderOp::unsupported};
static constexpr auto kEqualOps = TypedOps{BuilderOp::cmpeq_n_floats,
BuilderOp::cmpeq_n_ints,
BuilderOp::cmpeq_n_ints,
BuilderOp::cmpeq_n_ints};
static constexpr auto kNotEqualOps = TypedOps{BuilderOp::cmpne_n_floats,
BuilderOp::cmpne_n_ints,
BuilderOp::cmpne_n_ints,
BuilderOp::cmpne_n_ints};
static constexpr auto kMinOps = TypedOps{BuilderOp::min_n_floats,
BuilderOp::min_n_ints,
BuilderOp::min_n_uints,
BuilderOp::min_n_uints};
static constexpr auto kMaxOps = TypedOps{BuilderOp::max_n_floats,
BuilderOp::max_n_ints,
BuilderOp::max_n_uints,
BuilderOp::max_n_uints};
static constexpr auto kMixOps = TypedOps{BuilderOp::mix_n_floats,
BuilderOp::unsupported,
BuilderOp::unsupported,
BuilderOp::unsupported};
};
class LValue {
public:
virtual ~LValue() = default;
/**
* Returns an LValue for the passed-in expression; if the expression isn't supported as an
* LValue, returns nullptr.
*/
static std::unique_ptr<LValue> Make(const Expression& e);
/** Copies the top-of-stack value into this lvalue, without discarding it from the stack. */
void store(Generator* gen);
/** Pushes the lvalue onto the top-of-stack. */
void push(Generator* gen);
/**
* Returns the value slots associated with this LValue. For instance, a plain four-slot Variable
* will have monotonically increasing slots like {5,6,7,8}.
*/
struct SlotMap {
SkTArray<int> slots; // the destination slots
};
virtual SlotMap getSlotMap(Generator* gen) const = 0;
/** Returns true if this lvalue actually refers to a uniform. */
virtual bool isUniform() const = 0;
};
class VariableLValue final : public LValue {
public:
VariableLValue(const Variable* v) : fVariable(v) {}
SlotMap getSlotMap(Generator* gen) const override {
// Map every slot in the variable, in consecutive order, e.g. a half4 at slot 5 = {5,6,7,8}.
SlotMap out;
SlotRange range = this->isUniform() ? gen->getUniformSlots(*fVariable)
: gen->getVariableSlots(*fVariable);
out.slots.resize(range.count);
std::iota(out.slots.begin(), out.slots.end(), range.index);
return out;
}
bool isUniform() const override {
return Generator::IsUniform(*fVariable);
}
const Variable* fVariable;
};
class SwizzleLValue final : public LValue {
public:
SwizzleLValue(std::unique_ptr<LValue> p, const ComponentArray& c)
: fParent(std::move(p))
, fComponents(c) {}
SlotMap getSlotMap(Generator* gen) const override {
// Get slots from the parent expression.
SlotMap in = fParent->getSlotMap(gen);
// Rearrange the slots based to honor the swizzle components.
SlotMap out;
out.slots.resize(fComponents.size());
for (int index = 0; index < fComponents.size(); ++index) {
SkASSERT(fComponents[index] < in.slots.size());
out.slots[index] = in.slots[fComponents[index]];
}
return out;
}
bool isUniform() const override {
return fParent->isUniform();
}
std::unique_ptr<LValue> fParent;
const ComponentArray& fComponents;
};
class IndexLValue final : public LValue {
public:
IndexLValue(std::unique_ptr<LValue> p, const Expression& i, const Type& ti)
: fParent(std::move(p))
, fIndexExpr(i)
, fIndexedType(ti) {}
SlotMap getSlotMap(Generator* gen) const override {
// Get slots from the parent expression.
SlotMap in = fParent->getSlotMap(gen);
// Take a subset of the parent's slots.
// TODO(skia:13676): support non-constant indices
int numElements = fIndexedType.slotCount();
int startingIndex = numElements * fIndexExpr.as<Literal>().intValue();
SlotMap out;
out.slots.push_back_n(numElements, &in.slots[startingIndex]);
return out;
}
bool isUniform() const override {
return fParent->isUniform();
}
std::unique_ptr<LValue> fParent;
const Expression& fIndexExpr;
const Type& fIndexedType;
};
class FieldLValue final : public LValue {
public:
FieldLValue(std::unique_ptr<LValue> p, const FieldAccess& fieldAccess)
: fParent(std::move(p)) {
// Calculate the slot range of this field in the parent.
SkSpan<const Type::Field> fields = fieldAccess.base()->type().fields();
const int fieldIndex = fieldAccess.fieldIndex();
fInitialSlot = 0;
for (int index = 0; index < fieldIndex; ++index) {
fInitialSlot += fields[index].fType->slotCount();
}
fNumSlots = fields[fieldIndex].fType->slotCount();
}
SlotMap getSlotMap(Generator* gen) const override {
// Get the slot map from the base expression.
SlotMap in = fParent->getSlotMap(gen);
// Take a subset of the parent's slots.
SlotMap out;
out.slots.push_back_n(fNumSlots, &in.slots[fInitialSlot]);
return out;
}
bool isUniform() const override {
return fParent->isUniform();
}
std::unique_ptr<LValue> fParent;
int fInitialSlot = 0;
int fNumSlots = 0;
};
std::unique_ptr<LValue> LValue::Make(const Expression& e) {
if (e.is<VariableReference>()) {
return std::make_unique<VariableLValue>(e.as<VariableReference>().variable());
}
if (e.is<Swizzle>()) {
if (std::unique_ptr<LValue> base = LValue::Make(*e.as<Swizzle>().base())) {
return std::make_unique<SwizzleLValue>(std::move(base), e.as<Swizzle>().components());
}
}
if (e.is<FieldAccess>()) {
if (std::unique_ptr<LValue> base = LValue::Make(*e.as<FieldAccess>().base())) {
return std::make_unique<FieldLValue>(std::move(base), e.as<FieldAccess>());
}
}
if (e.is<IndexExpression>()) {
const IndexExpression& indexExpr = e.as<IndexExpression>();
// TODO(skia:13676): support non-constant indices
if (indexExpr.index()->is<Literal>()) {
if (std::unique_ptr<LValue> base = LValue::Make(*indexExpr.base())) {
return std::make_unique<IndexLValue>(std::move(base),
*indexExpr.index(),
indexExpr.type());
}
}
}
// TODO(skia:13676): add support for other kinds of lvalues
return nullptr;
}
void LValue::store(Generator* gen) {
SkASSERT(!this->isUniform());
SlotMap out = this->getSlotMap(gen);
// Copy our slots from the stack into their slots. The Builder will coalesce single-slot pushes
// into contiguous ranges where possible.
int offsetFromStackTop = out.slots.size();
for (Slot s : out.slots) {
gen->builder()->copy_stack_to_slots(SlotRange{s, 1}, offsetFromStackTop);
--offsetFromStackTop;
}
SkASSERT(offsetFromStackTop == 0);
}
void LValue::push(Generator* gen) {
SlotMap out = this->getSlotMap(gen);
// Push our slots onto the stack. The Builder will coalesce single-slot pushes into contiguous
// ranges where possible.
bool isUniform = this->isUniform();
for (Slot s : out.slots) {
if (isUniform) {
gen->builder()->push_uniform(SlotRange{s, 1});
} else {
gen->builder()->push_slots(SlotRange{s, 1});
}
}
}
static bool unsupported() {
// If MakeRasterPipelineProgram returns false, set a breakpoint here for more information.
return false;
}
void SlotManager::addSlotDebugInfoForGroup(const std::string& varName,
const Type& type,
Position pos,
int* groupIndex,
bool isFunctionReturnValue) {
SkASSERT(fSlotDebugInfo);
switch (type.typeKind()) {
case Type::TypeKind::kArray: {
int nslots = type.columns();
const Type& elemType = type.componentType();
for (int slot = 0; slot < nslots; ++slot) {
this->addSlotDebugInfoForGroup(varName + "[" + std::to_string(slot) + "]", elemType,
pos, groupIndex, isFunctionReturnValue);
}
break;
}
case Type::TypeKind::kStruct: {
for (const Type::Field& field : type.fields()) {
this->addSlotDebugInfoForGroup(varName + "." + std::string(field.fName),
*field.fType, pos, groupIndex,
isFunctionReturnValue);
}
break;
}
default:
SkASSERTF(0, "unsupported slot type %d", (int)type.typeKind());
[[fallthrough]];
case Type::TypeKind::kScalar:
case Type::TypeKind::kVector:
case Type::TypeKind::kMatrix: {
Type::NumberKind numberKind = type.componentType().numberKind();
int nslots = type.slotCount();
for (int slot = 0; slot < nslots; ++slot) {
SlotDebugInfo slotInfo;
slotInfo.name = varName;
slotInfo.columns = type.columns();
slotInfo.rows = type.rows();
slotInfo.componentIndex = slot;
slotInfo.groupIndex = (*groupIndex)++;
slotInfo.numberKind = numberKind;
slotInfo.pos = pos;
slotInfo.fnReturnValue = isFunctionReturnValue ? 1 : -1;
fSlotDebugInfo->push_back(std::move(slotInfo));
}
break;
}
}
}
void SlotManager::addSlotDebugInfo(const std::string& varName,
const Type& type,
Position pos,
bool isFunctionReturnValue) {
int groupIndex = 0;
this->addSlotDebugInfoForGroup(varName, type, pos, &groupIndex, isFunctionReturnValue);
SkASSERT((size_t)groupIndex == type.slotCount());
}
SlotRange SlotManager::createTemporarySlot(const Type& type) {
SkASSERT(type.slotCount() == 1);
// If we have an available slot to reclaim, take it now.
if (!fRecycledSlots.empty()) {
SlotRange result = {fRecycledSlots.back(), 1};
fRecycledSlots.pop_back();
return result;
}
// Synthesize a new temporary slot.
if (fSlotDebugInfo) {
// Our debug slot-info table should have the same length as the actual slot table.
SkASSERT(fSlotDebugInfo->size() == (size_t)fSlotCount);
// Add this temporary slot to the debug slot-info table. It's just scratch space which can
// be reused over the course of execution, so it doesn't get a name or type (uint will do).
this->addSlotDebugInfo(this->makeTempName(), type, Position{},
/*isFunctionReturnValue=*/false);
// Confirm that we added the expected number of slots.
SkASSERT(fSlotDebugInfo->size() == (size_t)(fSlotCount + 1));
}
SlotRange result = {fSlotCount, 1};
++fSlotCount;
return result;
}
void SlotManager::recycleTemporarySlot(SlotRange temporarySlot) {
SkASSERT(temporarySlot.count == 1);
fRecycledSlots.push_back(temporarySlot.index);
}
SlotRange SlotManager::createSlots(std::string name,
const Type& type,
Position pos,
bool isFunctionReturnValue) {
size_t nslots = type.slotCount();
if (nslots == 0) {
return {};
}
if (fSlotDebugInfo) {
// Our debug slot-info table should have the same length as the actual slot table.
SkASSERT(fSlotDebugInfo->size() == (size_t)fSlotCount);
// Append slot names and types to our debug slot-info table.
fSlotDebugInfo->reserve(fSlotCount + nslots);
this->addSlotDebugInfo(name, type, pos, isFunctionReturnValue);
// Confirm that we added the expected number of slots.
SkASSERT(fSlotDebugInfo->size() == (size_t)(fSlotCount + nslots));
}
SlotRange result = {fSlotCount, (int)nslots};
fSlotCount += nslots;
return result;
}
SlotRange SlotManager::getVariableSlots(const Variable& v) {
SlotRange* entry = fSlotMap.find(&v);
if (entry != nullptr) {
return *entry;
}
SlotRange range = this->createSlots(std::string(v.name()),
v.type(),
v.fPosition,
/*isFunctionReturnValue=*/false);
fSlotMap.set(&v, range);
return range;
}
SlotRange SlotManager::getFunctionSlots(const IRNode& callSite, const FunctionDeclaration& f) {
SlotRange* entry = fSlotMap.find(&callSite);
if (entry != nullptr) {
return *entry;
}
SlotRange range = this->createSlots("[" + std::string(f.name()) + "].result",
f.returnType(),
f.fPosition,
/*isFunctionReturnValue=*/true);
fSlotMap.set(&callSite, range);
return range;
}
int Generator::getFunctionDebugInfo(const FunctionDeclaration& decl) {
SkASSERT(fDebugTrace);
std::string name = decl.description();
// When generating the debug trace, we typically mark every function as `noinline`. This makes
// the trace more confusing, since this isn't in the source program, so remove it.
static constexpr std::string_view kNoInline = "noinline ";
if (skstd::starts_with(name, kNoInline)) {
name = name.substr(kNoInline.size());
}
// Look for a matching FunctionDebugInfo slot.
for (size_t index = 0; index < fDebugTrace->fFuncInfo.size(); ++index) {
if (fDebugTrace->fFuncInfo[index].name == name) {
return index;
}
}
// We've never called this function before; create a new slot to hold its information.
int slot = (int)fDebugTrace->fFuncInfo.size();
fDebugTrace->fFuncInfo.push_back(FunctionDebugInfo{std::move(name)});
return slot;
}
std::optional<SlotRange> Generator::writeFunction(const IRNode& callSite,
const FunctionDefinition& function) {
[[maybe_unused]] int funcIndex = -1;
if (fDebugTrace) {
funcIndex = this->getFunctionDebugInfo(function.declaration());
SkASSERT(funcIndex >= 0);
// TODO(debugger): add trace for function-enter
}
SlotRange lastFunctionResult = fCurrentFunctionResult;
fCurrentFunctionResult = this->getFunctionSlots(callSite, function.declaration());
if (!this->writeStatement(*function.body())) {
return std::nullopt;
}
SlotRange functionResult = fCurrentFunctionResult;
fCurrentFunctionResult = lastFunctionResult;
if (fDebugTrace) {
// TODO(debugger): add trace for function-exit
}
return functionResult;
}
bool Generator::writeGlobals() {
for (const ProgramElement* e : fProgram.elements()) {
if (e->is<GlobalVarDeclaration>()) {
const GlobalVarDeclaration& gvd = e->as<GlobalVarDeclaration>();
const VarDeclaration& decl = gvd.varDeclaration();
const Variable* var = decl.var();
if (var->type().isEffectChild()) {
// Associate each child effect variable with its numeric index.
SkASSERT(!fChildEffectMap.find(var));
int childEffectIndex = fChildEffectMap.count();
fChildEffectMap[var] = childEffectIndex;
continue;
}
// Opaque types include child processors and GL objects (samplers, textures, etc).
// Of those, only child processors are legal variables.
SkASSERT(!var->type().isVoid());
SkASSERT(!var->type().isOpaque());
// Builtin variables are system-defined, with special semantics.
if (int builtin = var->modifiers().fLayout.fBuiltin; builtin >= 0) {
if (builtin == SK_FRAGCOORD_BUILTIN) {
fBuilder.store_device_xy01(this->getVariableSlots(*var));
continue;
}
// The only builtin variable exposed to runtime effects is sk_FragCoord.
return unsupported();
}
if (IsUniform(*var)) {
// Create the uniform slot map in first-to-last order.
(void)this->getUniformSlots(*var);
continue;
}
// Other globals are treated as normal variable declarations.
if (!this->writeVarDeclaration(decl)) {
return unsupported();
}
}
}
return true;
}
bool Generator::writeStatement(const Statement& s) {
switch (s.kind()) {
case Statement::Kind::kBlock:
return this->writeBlock(s.as<Block>());
case Statement::Kind::kBreak:
return this->writeBreakStatement(s.as<BreakStatement>());
case Statement::Kind::kContinue:
return this->writeContinueStatement(s.as<ContinueStatement>());
case Statement::Kind::kDo:
return this->writeDoStatement(s.as<DoStatement>());
case Statement::Kind::kExpression:
return this->writeExpressionStatement(s.as<ExpressionStatement>());
case Statement::Kind::kFor:
return this->writeForStatement(s.as<ForStatement>());
case Statement::Kind::kIf:
return this->writeIfStatement(s.as<IfStatement>());
case Statement::Kind::kNop:
return true;
case Statement::Kind::kReturn:
return this->writeReturnStatement(s.as<ReturnStatement>());
case Statement::Kind::kVarDeclaration:
return this->writeVarDeclaration(s.as<VarDeclaration>());
default:
return unsupported();
}
}
bool Generator::writeBlock(const Block& b) {
for (const std::unique_ptr<Statement>& stmt : b.children()) {
if (!this->writeStatement(*stmt)) {
return unsupported();
}
}
return true;
}
bool Generator::writeBreakStatement(const BreakStatement&) {
fBuilder.mask_off_loop_mask();
return true;
}
bool Generator::writeContinueStatement(const ContinueStatement&) {
// This could be written as one hand-tuned RasterPipeline op, but for now, we reuse existing ops
// to assemble a continue op.
// Set any currently-executing lanes in the continue-mask to true via push-pop.
SkASSERT(fCurrentContinueMask.count == 1);
fBuilder.push_literal_i(~0);
this->popToSlotRange(fCurrentContinueMask);
// Disable any currently-executing lanes from the loop mask.
fBuilder.mask_off_loop_mask();
return true;
}
bool Generator::writeDoStatement(const DoStatement& d) {
// Save off the original loop mask.
fBuilder.enableExecutionMaskWrites();
fBuilder.push_loop_mask();
// If `continue` is used in the loop...
Analysis::LoopControlFlowInfo loopInfo = Analysis::GetLoopControlFlowInfo(*d.statement());
AutoContinueMask autoContinueMask;
if (loopInfo.fHasContinue) {
// ... create a temporary slot for continue-mask storage.
autoContinueMask.enable(&fProgramSlots, *fProgram.fContext->fTypes.fUInt,
&fCurrentContinueMask);
}
// Write the do-loop body.
int labelID = fBuilder.nextLabelID();
fBuilder.label(labelID);
autoContinueMask.enterLoopBody(fBuilder);
if (!this->writeStatement(*d.statement())) {
return false;
}
autoContinueMask.exitLoopBody(fBuilder);
// Emit the test-expression, in order to combine it with the loop mask.
if (!this->pushExpression(*d.test())) {
return false;
}
// Mask off any lanes in the loop mask where the test-expression is false; this breaks the loop.
// We don't use the test expression for anything else, so jettison it.
fBuilder.merge_loop_mask();
this->discardExpression(/*slots=*/1);
// If any lanes are still running, go back to the top and run the loop body again.
fBuilder.branch_if_any_active_lanes(labelID);
// Restore the loop mask.
fBuilder.pop_loop_mask();
fBuilder.disableExecutionMaskWrites();
return true;
}
bool Generator::writeMasklessForStatement(const ForStatement& f) {
SkASSERT(f.unrollInfo());
SkASSERT(f.unrollInfo()->fCount > 0);
SkASSERT(f.initializer());
SkASSERT(f.test());
SkASSERT(f.next());
// If no lanes are active, skip over the loop entirely. This guards against looping forever;
// with no lanes active, we wouldn't be able to write the loop variable back to its slot, so
// we'd never make forward progress.
int loopExitID = fBuilder.nextLabelID();
int loopBodyID = fBuilder.nextLabelID();
fBuilder.branch_if_no_active_lanes(loopExitID);
// Run the loop initializer.
if (!this->writeStatement(*f.initializer())) {
return unsupported();
}
// Write the for-loop body. We know the for-loop has a standard ES2 unrollable structure, and
// that it runs for at least one iteration, so we can plow straight ahead into the loop body
// instead of running the loop-test first.
fBuilder.label(loopBodyID);
if (!this->writeStatement(*f.statement())) {
return unsupported();
}
// If the loop only runs for a single iteration, we are already done. If not...
if (f.unrollInfo()->fCount > 1) {
// ... run the next-expression, and immediately discard its result.
if (!this->pushExpression(*f.next(), /*usesResult=*/false)) {
return unsupported();
}
this->discardExpression(f.next()->type().slotCount());
// Run the test-expression, and repeat the loop until the test-expression evaluates false.
if (!this->pushExpression(*f.test())) {
return unsupported();
}
fBuilder.branch_if_no_active_lanes_on_stack_top_equal(0, loopBodyID);
// Jettison the test-expression.
this->discardExpression(/*slots=*/1);
}
fBuilder.label(loopExitID);
return true;
}
bool Generator::writeForStatement(const ForStatement& f) {
// If we've determined that the loop does not run, omit its code entirely.
if (f.unrollInfo() && f.unrollInfo()->fCount == 0) {
return true;
}
// If the loop doesn't escape early due to a `continue`, `break` or `return`, and the loop
// conforms to ES2 structure, we know that we will run the full number of iterations across all
// lanes and don't need to use a loop mask.
Analysis::LoopControlFlowInfo loopInfo = Analysis::GetLoopControlFlowInfo(*f.statement());
if (!loopInfo.fHasContinue && !loopInfo.fHasBreak && !loopInfo.fHasReturn && f.unrollInfo()) {
return this->writeMasklessForStatement(f);
}
// Run the loop initializer.
if (f.initializer() && !this->writeStatement(*f.initializer())) {
return unsupported();
}
AutoContinueMask autoContinueMask;
if (loopInfo.fHasContinue) {
// Acquire a temporary slot for continue-mask storage.
autoContinueMask.enable(&fProgramSlots, *fProgram.fContext->fTypes.fUInt,
&fCurrentContinueMask);
}
// Save off the original loop mask.
fBuilder.enableExecutionMaskWrites();
fBuilder.push_loop_mask();
int loopTestID = fBuilder.nextLabelID();
int loopBodyID = fBuilder.nextLabelID();
// Jump down to the loop test so we can fall out of the loop immediately if it's zero-iteration.
fBuilder.jump(loopTestID);
// Write the for-loop body.
fBuilder.label(loopBodyID);
autoContinueMask.enterLoopBody(fBuilder);
if (!this->writeStatement(*f.statement())) {
return unsupported();
}
autoContinueMask.exitLoopBody(fBuilder);
// Run the next-expression. Immediately discard its result.
if (f.next()) {
if (!this->pushExpression(*f.next(), /*usesResult=*/false)) {
return unsupported();
}
this->discardExpression(f.next()->type().slotCount());
}
fBuilder.label(loopTestID);
if (f.test()) {
// Emit the test-expression, in order to combine it with the loop mask.
if (!this->pushExpression(*f.test())) {
return unsupported();
}
// Mask off any lanes in the loop mask where the test-expression is false; this breaks the
// loop. We don't use the test expression for anything else, so jettison it.
fBuilder.merge_loop_mask();
this->discardExpression(/*slots=*/1);
}
// If any lanes are still running, go back to the top and run the loop body again.
fBuilder.branch_if_any_active_lanes(loopBodyID);
// Restore the loop mask.
fBuilder.pop_loop_mask();
fBuilder.disableExecutionMaskWrites();
return true;
}
bool Generator::writeExpressionStatement(const ExpressionStatement& e) {
if (!this->pushExpression(*e.expression(), /*usesResult=*/false)) {
return unsupported();
}
this->discardExpression(e.expression()->type().slotCount());
return true;
}
bool Generator::writeDynamicallyUniformIfStatement(const IfStatement& i) {
SkASSERT(Analysis::IsDynamicallyUniformExpression(*i.test()));
int falseLabelID = fBuilder.nextLabelID();
int exitLabelID = fBuilder.nextLabelID();
if (!this->pushExpression(*i.test())) {
return unsupported();
}
fBuilder.branch_if_no_active_lanes_on_stack_top_equal(~0, falseLabelID);
if (!this->writeStatement(*i.ifTrue())) {
return unsupported();
}
if (!i.ifFalse()) {
// We don't have an if-false condition at all.
fBuilder.label(falseLabelID);
} else {
// We do have an if-false condition. We've just completed the if-true block, so we need to
// jump past the if-false block to avoid executing it.
fBuilder.jump(exitLabelID);
// The if-false block starts here.
fBuilder.label(falseLabelID);
if (!this->writeStatement(*i.ifFalse())) {
return unsupported();
}
fBuilder.label(exitLabelID);
}
// Jettison the test-expression.
this->discardExpression(/*slots=*/1);
return true;
}
bool Generator::writeIfStatement(const IfStatement& i) {
// If the test condition is known to be uniform, we can skip over the untrue portion entirely.
if (Analysis::IsDynamicallyUniformExpression(*i.test())) {
return this->writeDynamicallyUniformIfStatement(i);
}
// Save the current condition-mask.
fBuilder.enableExecutionMaskWrites();
fBuilder.push_condition_mask();
// Push the test condition mask.
if (!this->pushExpression(*i.test())) {
return unsupported();
}
// Merge the current condition-mask with the test condition, then run the if-true branch.
fBuilder.merge_condition_mask();
if (!this->writeStatement(*i.ifTrue())) {
return unsupported();
}
if (i.ifFalse()) {
// Negate the test-condition, then reapply it to the condition-mask.
// Then, run the if-false branch.
fBuilder.unary_op(BuilderOp::bitwise_not_int, /*slots=*/1);
fBuilder.merge_condition_mask();
if (!this->writeStatement(*i.ifFalse())) {
return unsupported();
}
}
// Jettison the test-expression, and restore the the condition-mask.
this->discardExpression(/*slots=*/1);
fBuilder.pop_condition_mask();
fBuilder.disableExecutionMaskWrites();
return true;
}
bool Generator::writeReturnStatement(const ReturnStatement& r) {
if (r.expression()) {
if (!this->pushExpression(*r.expression())) {
return unsupported();
}
this->popToSlotRange(fCurrentFunctionResult);
}
if (fBuilder.executionMaskWritesAreEnabled() && this->needsReturnMask()) {
fBuilder.mask_off_return_mask();
}
return true;
}
bool Generator::writeVarDeclaration(const VarDeclaration& v) {
if (v.value()) {
if (!this->pushExpression(*v.value())) {
return unsupported();
}
this->popToSlotRangeUnmasked(this->getVariableSlots(*v.var()));
} else {
this->zeroSlotRangeUnmasked(this->getVariableSlots(*v.var()));
}
return true;
}
bool Generator::pushExpression(const Expression& e, bool usesResult) {
switch (e.kind()) {
case Expression::Kind::kBinary:
return this->pushBinaryExpression(e.as<BinaryExpression>());
case Expression::Kind::kChildCall:
return this->pushChildCall(e.as<ChildCall>());
case Expression::Kind::kConstructorCompound:
return this->pushConstructorCompound(e.as<ConstructorCompound>());
case Expression::Kind::kConstructorCompoundCast:
case Expression::Kind::kConstructorScalarCast:
return this->pushConstructorCast(e.asAnyConstructor());
case Expression::Kind::kConstructorDiagonalMatrix:
return this->pushConstructorDiagonalMatrix(e.as<ConstructorDiagonalMatrix>());
case Expression::Kind::kConstructorMatrixResize:
return this->pushConstructorMatrixResize(e.as<ConstructorMatrixResize>());
case Expression::Kind::kConstructorSplat:
return this->pushConstructorSplat(e.as<ConstructorSplat>());
case Expression::Kind::kFunctionCall:
return this->pushFunctionCall(e.as<FunctionCall>());
case Expression::Kind::kIndex:
return this->pushIndexExpression(e.as<IndexExpression>());
case Expression::Kind::kLiteral:
return this->pushLiteral(e.as<Literal>());
case Expression::Kind::kPrefix:
return this->pushPrefixExpression(e.as<PrefixExpression>());
case Expression::Kind::kPostfix:
return this->pushPostfixExpression(e.as<PostfixExpression>(), usesResult);
case Expression::Kind::kSwizzle:
return this->pushSwizzle(e.as<Swizzle>());
case Expression::Kind::kTernary:
return this->pushTernaryExpression(e.as<TernaryExpression>());
case Expression::Kind::kVariableReference:
return this->pushVariableReference(e.as<VariableReference>());
default:
return unsupported();
}
}
BuilderOp Generator::GetTypedOp(const SkSL::Type& type, const TypedOps& ops) {
switch (type.componentType().numberKind()) {
case Type::NumberKind::kFloat: return ops.fFloatOp;
case Type::NumberKind::kSigned: return ops.fSignedOp;
case Type::NumberKind::kUnsigned: return ops.fUnsignedOp;
case Type::NumberKind::kBoolean: return ops.fBooleanOp;
default: return BuilderOp::unsupported;
}
}
bool Generator::unaryOp(const SkSL::Type& type, const TypedOps& ops) {
BuilderOp op = GetTypedOp(type, ops);
if (op == BuilderOp::unsupported) {
return unsupported();
}
fBuilder.unary_op(op, type.slotCount());
return true;
}
bool Generator::binaryOp(const SkSL::Type& type, const TypedOps& ops) {
BuilderOp op = GetTypedOp(type, ops);
if (op == BuilderOp::unsupported) {
return unsupported();
}
fBuilder.binary_op(op, type.slotCount());
return true;
}
bool Generator::ternaryOp(const SkSL::Type& type, const TypedOps& ops) {
BuilderOp op = GetTypedOp(type, ops);
if (op == BuilderOp::unsupported) {
return unsupported();
}
fBuilder.ternary_op(op, type.slotCount());
return true;
}
void Generator::foldWithMultiOp(BuilderOp op, int elements) {
// Fold the top N elements on the stack using an op that supports multiple slots, e.g.:
// (A + B + C + D) -> add_2_floats $0..1 += $2..3
// add_float $0 += $1
for (; elements >= 8; elements -= 4) {
fBuilder.binary_op(op, /*slots=*/4);
}
for (; elements >= 6; elements -= 3) {
fBuilder.binary_op(op, /*slots=*/3);
}
for (; elements >= 4; elements -= 2) {
fBuilder.binary_op(op, /*slots=*/2);
}
for (; elements >= 2; elements -= 1) {
fBuilder.binary_op(op, /*slots=*/1);
}
}
bool Generator::pushLValueOrExpression(LValue* lvalue, const Expression& expr) {
if (lvalue) {
lvalue->push(this);
return true;
}
return this->pushExpression(expr);
}
bool Generator::pushMatrixMultiply(LValue* lvalue,
const Expression& left,
const Expression& right,
int leftColumns,
int leftRows,
int rightColumns,
int rightRows) {
SkASSERT(left.type().isMatrix() || left.type().isVector());
SkASSERT(right.type().isMatrix() || right.type().isVector());
SkASSERT(leftColumns == rightRows);
int outColumns = rightColumns,
outRows = leftRows;
// Push the left matrix onto the adjacent-neighbor stack. We transpose it so that we can copy
// rows from it in a single op, instead of gathering one element at a time.
this->nextTempStack();
if (!this->pushLValueOrExpression(lvalue, left)) {
return unsupported();
}
fBuilder.transpose(leftColumns, leftRows);
// Push the right matrix as well, then go back to the primary stack.
if (!this->pushExpression(right)) {
return unsupported();
}
this->previousTempStack();
// Calculate the offsets of the left- and right-matrix, relative to the stack-top.
int leftMtxBase = left.type().slotCount() + right.type().slotCount() - leftColumns;
int rightMtxBase = right.type().slotCount() - leftColumns;
// Emit each matrix element.
for (int c = 0; c < outColumns; ++c) {
for (int r = 0; r < outRows; ++r) {
// Dot a vector from left[*][r] with right[c][*].
// (Because the left=matrix has been transposed, we actually pull left[r][*], which
// allows us to clone a column at once instead of cloning each slot individually.)
this->pushCloneFromNextTempStack(leftColumns, leftMtxBase - r * leftColumns);
this->pushCloneFromNextTempStack(leftColumns, rightMtxBase - c * leftColumns);
fBuilder.dot_floats(leftColumns);
}
}
// Dispose of the source matrices on the adjacent-neighbor stack.
this->nextTempStack();
this->discardExpression(left.type().slotCount());
this->discardExpression(right.type().slotCount());
this->previousTempStack();
// If this multiply was actually an assignment (via *=), write the result back to the lvalue.
if (lvalue) {
lvalue->store(this);
}
return true;
}
bool Generator::pushBinaryExpression(const BinaryExpression& e) {
return this->pushBinaryExpression(*e.left(), e.getOperator(), *e.right());
}
bool Generator::pushBinaryExpression(const Expression& left, Operator op, const Expression& right) {
// Rewrite greater-than ops as their less-than equivalents.
if (op.kind() == OperatorKind::GT) {
return this->pushBinaryExpression(right, OperatorKind::LT, left);
}
if (op.kind() == OperatorKind::GTEQ) {
return this->pushBinaryExpression(right, OperatorKind::LTEQ, left);
}
// Emit comma expressions.
if (op.kind() == OperatorKind::COMMA) {
if (Analysis::HasSideEffects(left)) {
if (!this->pushExpression(left, /*usesResult=*/false)) {
return unsupported();
}
this->discardExpression(left.type().slotCount());
}
return this->pushExpression(right);
}
// Handle binary expressions with mismatched types.
bool vectorizeLeft = false, vectorizeRight = false;
if (!left.type().matches(right.type())) {
if (left.type().componentType().numberKind() != right.type().componentType().numberKind()) {
return unsupported();
}
if (left.type().isScalar() && (right.type().isVector() || right.type().isMatrix())) {
vectorizeLeft = true;
} else if ((left.type().isVector() || left.type().isMatrix()) && right.type().isScalar()) {
vectorizeRight = true;
}
}
const Type& type = vectorizeLeft ? right.type() : left.type();
// If this is an assignment...
std::unique_ptr<LValue> lvalue;
if (op.isAssignment()) {
// ... turn the left side into an lvalue.
lvalue = LValue::Make(left);
if (!lvalue) {
return unsupported();
}
// Handle simple assignment (`var = expr`).
if (op.kind() == OperatorKind::EQ) {
if (!this->pushExpression(right)) {
return unsupported();
}
lvalue->store(this);
return true;
}
// Strip off the assignment from the op (turning += into +).
op = op.removeAssignment();
}
// Handle matrix multiplication (MxM/MxV/VxM).
if (op.kind() == OperatorKind::STAR) {
// Matrix * matrix:
if (left.type().isMatrix() && right.type().isMatrix()) {
return this->pushMatrixMultiply(lvalue.get(), left, right,
left.type().columns(), left.type().rows(),
right.type().columns(), right.type().rows());
}
// Vector * matrix:
if (left.type().isVector() && right.type().isMatrix()) {
return this->pushMatrixMultiply(lvalue.get(), left, right,
left.type().columns(), 1,
right.type().columns(), right.type().rows());
}
// Matrix * vector:
if (left.type().isMatrix() && right.type().isVector()) {
return this->pushMatrixMultiply(lvalue.get(), left, right,
left.type().columns(), left.type().rows(),
1, right.type().columns());
}
}
if (!vectorizeLeft && !vectorizeRight && !type.matches(right.type())) {
// We have mismatched types but don't know how to handle them.
return unsupported();
}
// Handle binary ops which require short-circuiting.
switch (op.kind()) {
case OperatorKind::LOGICALAND:
if (Analysis::HasSideEffects(right)) {
// If the RHS has side effects, we rewrite `a && b` as `a ? b : false`. This
// generates pretty solid code and gives us the required short-circuit behavior.
SkASSERT(!op.isAssignment());
SkASSERT(type.componentType().isBoolean());
SkASSERT(type.slotCount() == 1); // operator&& only works with scalar types
Literal falseLiteral{Position{}, 0.0, &right.type()};
return this->pushTernaryExpression(left, right, falseLiteral);
}
break;
case OperatorKind::LOGICALOR:
if (Analysis::HasSideEffects(right)) {
// If the RHS has side effects, we rewrite `a || b` as `a ? true : b`.
SkASSERT(!op.isAssignment());
SkASSERT(type.componentType().isBoolean());
SkASSERT(type.slotCount() == 1); // operator|| only works with scalar types
Literal trueLiteral{Position{}, 1.0, &right.type()};
return this->pushTernaryExpression(left, trueLiteral, right);
}
break;
default:
break;
}
// Push the left- and right-expressions onto the stack.
if (!this->pushLValueOrExpression(lvalue.get(), left)) {
return unsupported();
}
if (vectorizeLeft) {
fBuilder.push_duplicates(right.type().slotCount() - 1);
}
if (!this->pushExpression(right)) {
return unsupported();
}
if (vectorizeRight) {
fBuilder.push_duplicates(left.type().slotCount() - 1);
}
switch (op.kind()) {
case OperatorKind::PLUS:
if (!this->binaryOp(type, kAddOps)) {
return unsupported();
}
break;
case OperatorKind::MINUS:
if (!this->binaryOp(type, kSubtractOps)) {
return unsupported();
}
break;
case OperatorKind::STAR:
if (!this->binaryOp(type, kMultiplyOps)) {
return unsupported();
}
break;
case OperatorKind::SLASH:
if (!this->binaryOp(type, kDivideOps)) {
return unsupported();
}
break;
case OperatorKind::LT:
case OperatorKind::GT:
if (!this->binaryOp(type, kLessThanOps)) {
return unsupported();
}
SkASSERT(type.slotCount() == 1); // operator< only works with scalar types
break;
case OperatorKind::LTEQ:
case OperatorKind::GTEQ:
if (!this->binaryOp(type, kLessThanEqualOps)) {
return unsupported();
}
SkASSERT(type.slotCount() == 1); // operator<= only works with scalar types
break;
case OperatorKind::EQEQ:
if (!this->binaryOp(type, kEqualOps)) {
return unsupported();
}
// equal(x,y) returns a vector; use & to fold into a scalar.
this->foldWithMultiOp(BuilderOp::bitwise_and_n_ints, type.slotCount());
break;
case OperatorKind::NEQ:
if (!this->binaryOp(type, kNotEqualOps)) {
return unsupported();
}
// notEqual(x,y) returns a vector; use | to fold into a scalar.
this->foldWithMultiOp(BuilderOp::bitwise_or_n_ints, type.slotCount());
break;
case OperatorKind::LOGICALAND:
case OperatorKind::BITWISEAND:
// For logical-and, we verified above that the RHS does not have side effects, so we
// don't need to worry about short-circuiting side effects.
fBuilder.binary_op(BuilderOp::bitwise_and_n_ints, type.slotCount());
break;
case OperatorKind::LOGICALOR:
case OperatorKind::BITWISEOR:
// For logical-or, we verified above that the RHS does not have side effects.
fBuilder.binary_op(BuilderOp::bitwise_or_n_ints, type.slotCount());
break;
case OperatorKind::LOGICALXOR:
case OperatorKind::BITWISEXOR:
// Logical-xor does not short circuit.
fBuilder.binary_op(BuilderOp::bitwise_xor_n_ints, type.slotCount());
break;
default:
return unsupported();
}
// If we have an lvalue, we need to write the result back into it.
if (lvalue) {
lvalue->store(this);
}
return true;
}
bool Generator::pushConstructorCompound(const ConstructorCompound& c) {
for (const std::unique_ptr<Expression> &arg : c.arguments()) {
if (!this->pushExpression(*arg)) {
return unsupported();
}
}
return true;
}
bool Generator::pushChildCall(const ChildCall& c) {
int* childIdx = fChildEffectMap.find(&c.child());
SkASSERT(childIdx != nullptr);
SkASSERT(c.arguments().size() >= 1);
// Save the dst.rgba fields; these hold our execution masks, and could potentially be
// clobbered by the child effect.
fBuilder.push_dst_rgba();
// All child calls have at least one argument.
const Expression* arg = c.arguments()[0].get();
if (!this->pushExpression(*arg)) {
return unsupported();
}
// Copy arguments from the stack into src/dst as required by this particular child-call.
switch (c.child().type().typeKind()) {
case Type::TypeKind::kShader: {
// The argument must be a float2.
SkASSERT(c.arguments().size() == 1);
SkASSERT(arg->type().matches(*fProgram.fContext->fTypes.fFloat2));
fBuilder.pop_src_rg();
fBuilder.invoke_shader(*childIdx);
break;
}
case Type::TypeKind::kColorFilter: {
// The argument must be a half4/float4.
SkASSERT(c.arguments().size() == 1);
SkASSERT(arg->type().matches(*fProgram.fContext->fTypes.fHalf4) ||
arg->type().matches(*fProgram.fContext->fTypes.fFloat4));
fBuilder.pop_src_rgba();
fBuilder.invoke_color_filter(*childIdx);
break;
}
case Type::TypeKind::kBlender: {
// The first argument must be a half4/float4.
SkASSERT(c.arguments().size() == 2);
SkASSERT(arg->type().matches(*fProgram.fContext->fTypes.fHalf4) ||
arg->type().matches(*fProgram.fContext->fTypes.fFloat4));
// The second argument must also be a half4/float4.
arg = c.arguments()[1].get();
SkASSERT(arg->type().matches(*fProgram.fContext->fTypes.fHalf4) ||
arg->type().matches(*fProgram.fContext->fTypes.fFloat4));
if (!this->pushExpression(*arg)) {
return unsupported();
}
fBuilder.pop_dst_rgba();
fBuilder.pop_src_rgba();
fBuilder.invoke_blender(*childIdx);
break;
}
default: {
SkDEBUGFAILF("cannot sample from type '%s'", c.child().type().description().c_str());
}
}
// Restore dst.rgba so our execution masks are back to normal.
fBuilder.pop_dst_rgba();
// The child call has returned the result color via src.rgba; push it onto the stack.
fBuilder.push_src_rgba();
return true;
}
bool Generator::pushConstructorCast(const AnyConstructor& c) {
SkASSERT(c.argumentSpan().size() == 1);
const Expression& inner = *c.argumentSpan().front();
SkASSERT(inner.type().slotCount() == c.type().slotCount());
if (!this->pushExpression(inner)) {
return unsupported();
}
if (inner.type().componentType().numberKind() == c.type().componentType().numberKind()) {
// Since we ignore type precision, this cast is effectively a no-op.
return true;
}
if (inner.type().componentType().isSigned() && c.type().componentType().isUnsigned()) {
// Treat uint(int) as a no-op.
return true;
}
if (inner.type().componentType().isUnsigned() && c.type().componentType().isSigned()) {
// Treat int(uint) as a no-op.
return true;
}
if (c.type().componentType().isBoolean()) {
// Converting int or float to boolean can be accomplished via `notEqual(x, 0)`.
fBuilder.push_zeros(c.type().slotCount());
return this->binaryOp(inner.type(), kNotEqualOps);
}
if (inner.type().componentType().isBoolean()) {
// Converting boolean to int or float can be accomplished via bitwise-and.
if (c.type().componentType().isFloat()) {
fBuilder.push_literal_f(1.0f);
} else if (c.type().componentType().isSigned() || c.type().componentType().isUnsigned()) {
fBuilder.push_literal_i(1);
} else {
SkDEBUGFAILF("unexpected cast from bool to %s", c.type().description().c_str());
return unsupported();
}
fBuilder.push_duplicates(c.type().slotCount() - 1);
fBuilder.binary_op(BuilderOp::bitwise_and_n_ints, c.type().slotCount());
return true;
}
// We have dedicated ops to cast between float and integer types.
if (inner.type().componentType().isFloat()) {
if (c.type().componentType().isSigned()) {
fBuilder.unary_op(BuilderOp::cast_to_int_from_float, c.type().slotCount());
return true;
}
if (c.type().componentType().isUnsigned()) {
fBuilder.unary_op(BuilderOp::cast_to_uint_from_float, c.type().slotCount());
return true;
}
} else if (c.type().componentType().isFloat()) {
if (inner.type().componentType().isSigned()) {
fBuilder.unary_op(BuilderOp::cast_to_float_from_int, c.type().slotCount());
return true;
}
if (inner.type().componentType().isUnsigned()) {
fBuilder.unary_op(BuilderOp::cast_to_float_from_uint, c.type().slotCount());
return true;
}
}
SkDEBUGFAILF("unexpected cast from %s to %s",
c.type().description().c_str(), inner.type().description().c_str());
return unsupported();
}
bool Generator::pushConstructorDiagonalMatrix(const ConstructorDiagonalMatrix& c) {
fBuilder.push_zeros(1);
if (!this->pushExpression(*c.argument())) {
return unsupported();
}
fBuilder.diagonal_matrix(c.type().columns(), c.type().rows());
return true;
}
bool Generator::pushConstructorMatrixResize(const ConstructorMatrixResize& c) {
if (!this->pushExpression(*c.argument())) {
return unsupported();
}
fBuilder.matrix_resize(c.argument()->type().columns(),
c.argument()->type().rows(),
c.type().columns(),
c.type().rows());
return true;
}
bool Generator::pushConstructorSplat(const ConstructorSplat& c) {
if (!this->pushExpression(*c.argument())) {
return unsupported();
}
fBuilder.push_duplicates(c.type().slotCount() - 1);
return true;
}
bool Generator::pushFunctionCall(const FunctionCall& c) {
if (c.function().isIntrinsic()) {
return this->pushIntrinsic(c);
}
// Keep track of the current function.
const FunctionDefinition* lastFunction = fCurrentFunction;
fCurrentFunction = c.function().definition();
// Skip over the function body entirely if there are no active lanes.
// (If the function call was trivial, it would likely have been inlined in the frontend, so this
// is likely to save a significant amount of work if the lanes are all dead.)
int skipLabelID = fBuilder.nextLabelID();
fBuilder.branch_if_no_active_lanes(skipLabelID);
// Save off the return mask.
if (this->needsReturnMask()) {
fBuilder.enableExecutionMaskWrites();
fBuilder.push_return_mask();
}
// Write all the arguments into their parameter's variable slots. Because we never allow
// recursion, we don't need to worry about overwriting any existing values in those slots.
// (In fact, we don't even need to apply the write mask.)
SkTArray<std::unique_ptr<LValue>> lvalues;
lvalues.resize(c.arguments().size());
for (int index = 0; index < c.arguments().size(); ++index) {
const Expression& arg = *c.arguments()[index];
const Variable& param = *c.function().parameters()[index];
// Use LValues for out-parameters and inout-parameters, so we can store back to them later.
if (IsInoutParameter(param) || IsOutParameter(param)) {
lvalues[index] = LValue::Make(arg);
if (!lvalues[index]) {
return unsupported();
}
// There are no guarantees on the starting value of an out-parameter, so we only need to
// store the lvalues associated with an inout parameter.
if (IsInoutParameter(param)) {
lvalues[index]->push(this);
this->popToSlotRangeUnmasked(this->getVariableSlots(param));
}
} else {
// Copy input arguments into their respective parameter slots.
if (!this->pushExpression(arg)) {
return unsupported();
}
this->popToSlotRangeUnmasked(this->getVariableSlots(param));
}
}
// Emit the function body.
std::optional<SlotRange> r = this->writeFunction(c, *fCurrentFunction);
if (!r.has_value()) {
return unsupported();
}
// Restore the original return mask.
if (this->needsReturnMask()) {
fBuilder.pop_return_mask();
fBuilder.disableExecutionMaskWrites();
}
// We've returned back to the last function.
fCurrentFunction = lastFunction;
// Copy out-parameters and inout-parameters back to their homes.
for (int index = 0; index < c.arguments().size(); ++index) {
if (lvalues[index]) {
// Only out- and inout-parameters should have an associated lvalue.
const Variable& param = *c.function().parameters()[index];
SkASSERT(IsInoutParameter(param) || IsOutParameter(param));
// Copy the parameter's slots directly into the lvalue.
fBuilder.push_slots(this->getVariableSlots(param));
lvalues[index]->store(this);
this->discardExpression(param.type().slotCount());
}
}
// Copy the function result from its slots onto the stack.
fBuilder.push_slots(*r);
fBuilder.label(skipLabelID);
return true;
}
bool Generator::pushIndexExpression(const IndexExpression& i) {
std::unique_ptr<LValue> lvalue = LValue::Make(i);
if (!lvalue) {
return unsupported();
}
lvalue->push(this);
return true;
}
bool Generator::pushIntrinsic(const FunctionCall& c) {
const ExpressionArray& args = c.arguments();
switch (args.size()) {
case 1:
return this->pushIntrinsic(c.function().intrinsicKind(), *args[0]);
case 2:
return this->pushIntrinsic(c.function().intrinsicKind(), *args[0], *args[1]);
case 3:
return this->pushIntrinsic(c.function().intrinsicKind(), *args[0], *args[1], *args[2]);
default:
break;
}
return unsupported();
}
bool Generator::pushVectorizedExpression(const Expression& expr, const Type& vectorType) {
if (!this->pushExpression(expr)) {
return unsupported();
}
if (vectorType.slotCount() > expr.type().slotCount()) {
SkASSERT(expr.type().slotCount() == 1);
fBuilder.push_duplicates(vectorType.slotCount() - expr.type().slotCount());
}
return true;
}
bool Generator::pushIntrinsic(const TypedOps& ops, const Expression& arg0) {
if (!this->pushExpression(arg0)) {
return unsupported();
}
return this->unaryOp(arg0.type(), ops);
}
bool Generator::pushIntrinsic(BuilderOp builderOp, const Expression& arg0) {
if (!this->pushExpression(arg0)) {
return unsupported();
}
fBuilder.unary_op(builderOp, arg0.type().slotCount());
return true;
}
bool Generator::pushIntrinsic(IntrinsicKind intrinsic, const Expression& arg0) {
switch (intrinsic) {
case IntrinsicKind::k_abs_IntrinsicKind:
return this->pushIntrinsic(kAbsOps, arg0);
case IntrinsicKind::k_any_IntrinsicKind:
if (!this->pushExpression(arg0)) {
return unsupported();
}
this->foldWithMultiOp(BuilderOp::bitwise_or_n_ints, arg0.type().slotCount());
return true;
case IntrinsicKind::k_all_IntrinsicKind:
if (!this->pushExpression(arg0)) {
return unsupported();
}
this->foldWithMultiOp(BuilderOp::bitwise_and_n_ints, arg0.type().slotCount());
return true;
case IntrinsicKind::k_atan_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::atan_float, arg0);
case IntrinsicKind::k_ceil_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::ceil_float, arg0);
case IntrinsicKind::k_cos_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::cos_float, arg0);
case IntrinsicKind::k_degrees_IntrinsicKind: {
Literal lit180OverPi{Position{}, 57.2957795131f, &arg0.type().componentType()};
return this->pushBinaryExpression(arg0, OperatorKind::STAR, lit180OverPi);
}
case IntrinsicKind::k_floatBitsToInt_IntrinsicKind:
case IntrinsicKind::k_floatBitsToUint_IntrinsicKind:
case IntrinsicKind::k_intBitsToFloat_IntrinsicKind:
case IntrinsicKind::k_uintBitsToFloat_IntrinsicKind:
return this->pushExpression(arg0);
case IntrinsicKind::k_exp_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::exp_float, arg0);
case IntrinsicKind::k_floor_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::floor_float, arg0);
case IntrinsicKind::k_fract_IntrinsicKind:
// Implement fract as `x - floor(x)`.
if (!this->pushExpression(arg0)) {
return unsupported();
}
fBuilder.push_clone(arg0.type().slotCount());
fBuilder.unary_op(BuilderOp::floor_float, arg0.type().slotCount());
return this->binaryOp(arg0.type(), kSubtractOps);
case IntrinsicKind::k_length_IntrinsicKind:
if (!this->pushExpression(arg0)) {
return unsupported();
}
// Implement length as `sqrt(dot(x, x))`.
if (arg0.type().slotCount() > 1) {
fBuilder.push_clone(arg0.type().slotCount());
fBuilder.dot_floats(arg0.type().slotCount());
fBuilder.unary_op(BuilderOp::sqrt_float, 1);
} else {
// The length of a scalar is `sqrt(x^2)`, which is equivalent to `abs(x)`.
fBuilder.unary_op(BuilderOp::abs_float, 1);
}
return true;
case IntrinsicKind::k_not_IntrinsicKind:
return this->pushPrefixExpression(OperatorKind::LOGICALNOT, arg0);
case IntrinsicKind::k_radians_IntrinsicKind: {
Literal litPiOver180{Position{}, 0.01745329251f, &arg0.type().componentType()};
return this->pushBinaryExpression(arg0, OperatorKind::STAR, litPiOver180);
}
case IntrinsicKind::k_saturate_IntrinsicKind: {
// Implement saturate as clamp(arg, 0, 1).
Literal zeroLiteral{Position{}, 0.0, &arg0.type().componentType()};
Literal oneLiteral{Position{}, 1.0, &arg0.type().componentType()};
return this->pushIntrinsic(k_clamp_IntrinsicKind, arg0, zeroLiteral, oneLiteral);
}
case IntrinsicKind::k_sign_IntrinsicKind: {
// Implement floating-point sign() as `clamp(arg * FLT_MAX, -1, 1)`.
// FLT_MIN * FLT_MAX evaluates to 4, so multiplying any float value against FLT_MAX is
// sufficient to ensure that |value| is always 1 or greater (excluding zero and nan).
// Integer sign() doesn't need to worry about fractional values or nans, and can simply
// be `clamp(arg, -1, 1)`.
if (!this->pushExpression(arg0)) {
return unsupported();
}
if (arg0.type().componentType().isFloat()) {
Literal fltMaxLiteral{Position{}, FLT_MAX, &arg0.type().componentType()};
if (!this->pushVectorizedExpression(fltMaxLiteral, arg0.type())) {
return unsupported();
}
if (!this->binaryOp(arg0.type(), kMultiplyOps)) {
return unsupported();
}
}
Literal neg1Literal{Position{}, -1.0, &arg0.type().componentType()};
if (!this->pushVectorizedExpression(neg1Literal, arg0.type())) {
return unsupported();
}
if (!this->binaryOp(arg0.type(), kMaxOps)) {
return unsupported();
}
Literal pos1Literal{Position{}, 1.0, &arg0.type().componentType()};
if (!this->pushVectorizedExpression(pos1Literal, arg0.type())) {
return unsupported();
}
return this->binaryOp(arg0.type(), kMinOps);
}
case IntrinsicKind::k_sin_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::sin_float, arg0);
case IntrinsicKind::k_sqrt_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::sqrt_float, arg0);
case IntrinsicKind::k_tan_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::tan_float, arg0);
case IntrinsicKind::k_transpose_IntrinsicKind:
SkASSERT(arg0.type().isMatrix());
if (!this->pushExpression(arg0)) {
return unsupported();
}
fBuilder.transpose(arg0.type().columns(), arg0.type().rows());
return true;
default:
break;
}
return unsupported();
}
bool Generator::pushIntrinsic(const TypedOps& ops, const Expression& arg0, const Expression& arg1) {
if (!this->pushExpression(arg0) || !this->pushVectorizedExpression(arg1, arg0.type())) {
return unsupported();
}
return this->binaryOp(arg0.type(), ops);
}
bool Generator::pushIntrinsic(BuilderOp builderOp, const Expression& arg0, const Expression& arg1) {
if (!this->pushExpression(arg0) || !this->pushVectorizedExpression(arg1, arg0.type())) {
return unsupported();
}
fBuilder.binary_op(builderOp, arg0.type().slotCount());
return true;
}
bool Generator::pushIntrinsic(IntrinsicKind intrinsic,
const Expression& arg0,
const Expression& arg1) {
switch (intrinsic) {
case IntrinsicKind::k_atan_IntrinsicKind:
return this->pushIntrinsic(BuilderOp::atan2_n_floats, arg0, arg1);
case IntrinsicKind::k_cross_IntrinsicKind:
// Implement cross as `arg0.yzx * arg1.zxy - arg0.zxy * arg1.yzx`. We use two stacks so
// that each subexpression can be multiplied separately.
SkASSERT(arg0.type().matches(arg1.type()));
SkASSERT(arg0.type().slotCount() == 3);
SkASSERT(arg1.type().slotCount() == 3);
// Push `arg0.yzx` onto this stack and `arg0.zxy` onto the next stack.
if (!this->pushExpression(arg0)) {
return unsupported();
}
this->nextTempStack();
this->pushCloneFromPreviousTempStack(3);
fBuilder.swizzle(/*consumedSlots=*/3, {2, 0, 1});
this->previousTempStack();
fBuilder.swizzle(/*consumedSlots=*/3, {1, 2, 0});
// Push `arg1.zxy` onto this stack and `arg1.yzx` onto the next stack. Perform the
// multiply on each subexpression (`arg0.yzx * arg1.zxy` on the first stack, and
// `arg0.zxy * arg1.yzx` on the next).
if (!this->pushExpression(arg1)) {
return unsupported();
}
this->nextTempStack();
this->pushCloneFromPreviousTempStack(3);
fBuilder.swizzle(/*consumedSlots=*/3, {1, 2, 0});
fBuilder.binary_op(BuilderOp::mul_n_floats, 3);
this->previousTempStack();
fBuilder.swizzle(/*consumedSlots=*/3, {2, 0, 1});
fBuilder.binary_op(BuilderOp::mul_n_floats, 3);
// Migrate the result of the second subexpression (`arg0.zxy * arg1.yzx`) back onto the
// main stack and subtract it from the first subexpression (`arg0.yzx * arg1.zxy`).
this->pushCloneFromNextTempStack(3);
fBuilder.binary_op(BuilderOp::sub_n_floats, 3);
// Now that the calculation is complete, discard the subexpression on the next stack.
this->nextTempStack();
this->discardExpression(/*slots=*/3);
this->previousTempStack();
return true;
case IntrinsicKind::k_dot_IntrinsicKind:
// Implement dot as `a*b`, followed by folding via addition.
SkASSERT(arg0.type().matches(arg1.type()));
if (!this->pushExpression(arg0) || !this->pushExpression(arg1)) {
return unsupported();
}
fBuilder.dot_floats(arg0.type().slotCount());
return true;
case IntrinsicKind::k_equal_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kEqualOps, arg0, arg1);
case IntrinsicKind::k_notEqual_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kNotEqualOps, arg0, arg1);
case IntrinsicKind::k_lessThan_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kLessThanOps, arg0, arg1);
case IntrinsicKind::k_greaterThan_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kLessThanOps, arg1, arg0);
case IntrinsicKind::k_lessThanEqual_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kLessThanEqualOps, arg0, arg1);
case IntrinsicKind::k_greaterThanEqual_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kLessThanEqualOps, arg1, arg0);
case IntrinsicKind::k_min_IntrinsicKind:
SkASSERT(arg0.type().componentType().matches(arg1.type().componentType()));
return this->pushIntrinsic(kMinOps, arg0, arg1);
case IntrinsicKind::k_matrixCompMult_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(kMultiplyOps, arg0, arg1);
case IntrinsicKind::k_max_IntrinsicKind:
SkASSERT(arg0.type().componentType().matches(arg1.type().componentType()));
return this->pushIntrinsic(kMaxOps, arg0, arg1);
case IntrinsicKind::k_pow_IntrinsicKind:
SkASSERT(arg0.type().matches(arg1.type()));
return this->pushIntrinsic(BuilderOp::pow_n_floats, arg0, arg1);
case IntrinsicKind::k_step_IntrinsicKind: {
// Compute step as `float(lessThan(edge, x))`. We convert from boolean 0/~0 to floating
// point zero/one by using a bitwise-and against the bit-pattern of 1.0.
SkASSERT(arg0.type().componentType().matches(arg1.type().componentType()));
if (!this->pushVectorizedExpression(arg0, arg1.type()) || !this->pushExpression(arg1)) {
return unsupported();
}
if (!this->binaryOp(arg1.type(), kLessThanOps)) {
return unsupported();
}
Literal pos1Literal{Position{}, 1.0, &arg1.type().componentType()};
if (!this->pushVectorizedExpression(pos1Literal, arg1.type())) {
return unsupported();
}
fBuilder.binary_op(BuilderOp::bitwise_and_n_ints, arg1.type().slotCount());
return true;
}
default:
break;
}
return unsupported();
}
bool Generator::pushIntrinsic(IntrinsicKind intrinsic,
const Expression& arg0,
const Expression& arg1,
const Expression& arg2) {
switch (intrinsic) {
case IntrinsicKind::k_clamp_IntrinsicKind:
// Implement clamp as min(max(arg, low), high).
SkASSERT(arg0.type().componentType().matches(arg1.type().componentType()));
SkASSERT(arg0.type().componentType().matches(arg2.type().componentType()));
if (!this->pushExpression(arg0) || !this->pushVectorizedExpression(arg1, arg0.type())) {
return unsupported();
}
if (!this->binaryOp(arg0.type(), kMaxOps)) {
return unsupported();
}
if (!this->pushVectorizedExpression(arg2, arg0.type())) {
return unsupported();
}
if (!this->binaryOp(arg0.type(), kMinOps)) {
return unsupported();
}
return true;
case IntrinsicKind::k_mix_IntrinsicKind:
// TODO: implement mix(a,b,genBType)
if (!arg2.type().componentType().isFloat()) {
return unsupported();
}
SkASSERT(arg0.type().matches(arg1.type()));
SkASSERT(arg0.type().componentType().matches(arg2.type().componentType()));
if (!this->pushExpression(arg0) || !this->pushExpression(arg1)) {
return unsupported();
}
if (!this->pushVectorizedExpression(arg2, arg0.type())) {
return unsupported();
}
if (!this->ternaryOp(arg0.type(), kMixOps)) {
return unsupported();
}
return true;
default:
break;
}
return unsupported();
}
bool Generator::pushLiteral(const Literal& l) {
switch (l.type().numberKind()) {
case Type::NumberKind::kFloat:
fBuilder.push_literal_f(l.floatValue());
return true;
case Type::NumberKind::kSigned:
fBuilder.push_literal_i(l.intValue());
return true;
case Type::NumberKind::kUnsigned:
fBuilder.push_literal_u(l.intValue());
return true;
case Type::NumberKind::kBoolean:
fBuilder.push_literal_i(l.boolValue() ? ~0 : 0);
return true;
default:
SkUNREACHABLE;
}
}
bool Generator::pushPostfixExpression(const PostfixExpression& p, bool usesResult) {
// If the result is ignored...
if (!usesResult) {
// ... just emit a prefix expression instead.
return this->pushPrefixExpression(p.getOperator(), *p.operand());
}
// Get the operand as an lvalue, and push it onto the stack as-is.
std::unique_ptr<LValue> lvalue = LValue::Make(*p.operand());
if (!lvalue) {
return unsupported();
}
lvalue->push(this);
// Push a scratch copy of the operand.
fBuilder.push_clone(p.type().slotCount());
// Increment or decrement the scratch copy by one.
Literal oneLiteral{Position{}, 1.0, &p.type().componentType()};
if (!this->pushVectorizedExpression(oneLiteral, p.type())) {
return unsupported();
}
switch (p.getOperator().kind()) {
case OperatorKind::PLUSPLUS:
if (!this->binaryOp(p.type(), kAddOps)) {
return unsupported();
}
break;
case OperatorKind::MINUSMINUS:
if (!this->binaryOp(p.type(), kSubtractOps)) {
return unsupported();
}
break;
default:
SkUNREACHABLE;
}
// Write the new value back to the operand.
lvalue->store(this);
// Discard the scratch copy, leaving only the original value as-is.
this->discardExpression(p.type().slotCount());
return true;
}
bool Generator::pushPrefixExpression(const PrefixExpression& p) {
return this->pushPrefixExpression(p.getOperator(), *p.operand());
}
bool Generator::pushPrefixExpression(Operator op, const Expression& expr) {
switch (op.kind()) {
case OperatorKind::BITWISENOT:
case OperatorKind::LOGICALNOT:
// Handle operators ! and ~.
if (!this->pushExpression(expr)) {
return unsupported();
}
fBuilder.unary_op(BuilderOp::bitwise_not_int, expr.type().slotCount());
return true;
case OperatorKind::MINUS:
// Handle negation as a componentwise `0 - expr`.
fBuilder.push_zeros(expr.type().slotCount());
if (!this->pushExpression(expr)) {
return unsupported();
}
return this->binaryOp(expr.type(), kSubtractOps);
case OperatorKind::PLUSPLUS: {
// Rewrite as `expr += 1`.
Literal oneLiteral{Position{}, 1.0, &expr.type().componentType()};
return this->pushBinaryExpression(expr, OperatorKind::PLUSEQ, oneLiteral);
}
case OperatorKind::MINUSMINUS: {
// Rewrite as `expr -= 1`.
Literal oneLiteral{Position{}, 1.0, &expr.type().componentType()};
return this->pushBinaryExpression(expr, OperatorKind::MINUSEQ, oneLiteral);
}
default:
break;
}
return unsupported();
}
bool Generator::pushSwizzle(const Swizzle& s) {
SkASSERT(s.components().size() >= 1 && s.components().size() <= 4);
// Determine if the swizzle rearranges its elements, or if it's a simple subset of its elements.
// (A simple subset would be a sequential non-repeating range of components, like `.xyz` or
// `.yzw` or `.z`, but not `.xx` or `.xz`.)
bool isSimpleSubset = true;
for (int index = 1; index < s.components().size(); ++index) {
if (s.components()[index] != s.components()[0] + index) {
isSimpleSubset = false;
break;
}
}
// If this is a simple subset of a variable's slots...
if (isSimpleSubset && s.base()->is<VariableReference>()) {
// ... we can just push part of the variable directly onto the stack, rather than pushing
// the whole expression and then immediately cutting it down. (Either way works, but this
// saves a step.)
return this->pushVariableReferencePartial(
s.base()->as<VariableReference>(),
SlotRange{/*index=*/s.components()[0], /*count=*/s.components().size()});
}
// Push the base expression.
if (!this->pushExpression(*s.base())) {
return false;
}
// An identity swizzle doesn't rearrange the data; it just (potentially) discards tail elements.
if (isSimpleSubset && s.components()[0] == 0) {
int discardedElements = s.base()->type().slotCount() - s.components().size();
SkASSERT(discardedElements >= 0);
fBuilder.discard_stack(discardedElements);
return true;
}
// Perform the swizzle.
fBuilder.swizzle(s.base()->type().slotCount(), s.components());
return true;
}
bool Generator::pushTernaryExpression(const TernaryExpression& t) {
return this->pushTernaryExpression(*t.test(), *t.ifTrue(), *t.ifFalse());
}
bool Generator::pushDynamicallyUniformTernaryExpression(const Expression& test,
const Expression& ifTrue,
const Expression& ifFalse) {
SkASSERT(Analysis::IsDynamicallyUniformExpression(test));
int falseLabelID = fBuilder.nextLabelID();
int exitLabelID = fBuilder.nextLabelID();
// First, push the test-expression into a separate stack.
this->nextTempStack();
if (!this->pushExpression(test)) {
return unsupported();
}
// Branch to the true- or false-expression based on the test-expression. We can skip the
// non-true path entirely since the test is known to be uniform.
fBuilder.branch_if_no_active_lanes_on_stack_top_equal(~0, falseLabelID);
this->previousTempStack();
if (!this->pushExpression(ifTrue)) {
return unsupported();
}
fBuilder.jump(exitLabelID);
// The builder doesn't understand control flow, and assumes that every push moves the stack-top
// forwards. We need to manually balance out the `pushExpression` from the if-true path by
// moving the stack position backwards, so that the if-false path pushes its expression into the
// same as the if-true result.
this->discardExpression(/*slots=*/ifTrue.type().slotCount());
fBuilder.label(falseLabelID);
if (!this->pushExpression(ifFalse)) {
return unsupported();
}
fBuilder.label(exitLabelID);
// Jettison the text-expression from the separate stack.
this->nextTempStack();
this->discardExpression(/*slots=*/1);
this->previousTempStack();
return true;
}
bool Generator::pushTernaryExpression(const Expression& test,
const Expression& ifTrue,
const Expression& ifFalse) {
// If the test-expression is dynamically-uniform, we can skip over the non-true expressions
// entirely, and not need to involve the condition mask.
if (Analysis::IsDynamicallyUniformExpression(test)) {
return this->pushDynamicallyUniformTernaryExpression(test, ifTrue, ifFalse);
}
fBuilder.enableExecutionMaskWrites();
// First, push the current condition-mask and the test-expression into a separate stack.
this->nextTempStack();
fBuilder.push_condition_mask();
if (!this->pushExpression(test)) {
return unsupported();
}
this->previousTempStack();
// We can take some shortcuts with condition-mask handling if the false-expression is entirely
// side-effect free. (We can evaluate it without masking off its effects.) We always handle the
// condition mask properly for the test-expression and true-expression properly.
if (!Analysis::HasSideEffects(ifFalse)) {
// Push the false-expression onto the primary stack.
int cleanupLabelID = fBuilder.nextLabelID();
if (!this->pushExpression(ifFalse)) {
return unsupported();
}
// Next, merge the condition mask (on the separate stack) with the test expression.
this->nextTempStack();
fBuilder.merge_condition_mask();
this->previousTempStack();
// If no lanes are active, we can skip the true-expression entirely. This isn't super likely
// to happen, so it's probably only a win for non-trivial true-expressions.
if (!Analysis::IsTrivialExpression(ifTrue)) {
fBuilder.branch_if_no_active_lanes(cleanupLabelID);
}
// Push the true-expression onto the primary stack, immediately after the false-expression.
if (!this->pushExpression(ifTrue)) {
return unsupported();
}
// Use a select to conditionally mask-merge the true-expression and false-expression lanes.
fBuilder.select(/*slots=*/ifTrue.type().slotCount());
fBuilder.label(cleanupLabelID);
} else {
// Merge the condition mask (on the separate stack) with the test expression.
this->nextTempStack();
fBuilder.merge_condition_mask();
this->previousTempStack();
// Push the true-expression onto the primary stack.
if (!this->pushExpression(ifTrue)) {
return unsupported();
}
// Switch back to the test-expression stack temporarily, and negate the test condition.
this->nextTempStack();
fBuilder.unary_op(BuilderOp::bitwise_not_int, /*slots=*/1);
fBuilder.merge_condition_mask();
this->previousTempStack();
// Push the false-expression onto the primary stack, immediately after the true-expression.
if (!this->pushExpression(ifFalse)) {
return unsupported();
}
// Use a select to conditionally mask-merge the true-expression and false-expression lanes;
// the mask is already set up for this.
fBuilder.select(/*slots=*/ifTrue.type().slotCount());
}
// Restore the condition-mask to its original state and jettison the test-expression.
this->nextTempStack();
this->discardExpression(/*slots=*/1);
fBuilder.pop_condition_mask();
this->previousTempStack();
fBuilder.disableExecutionMaskWrites();
return true;
}
bool Generator::pushVariableReference(const VariableReference& v) {
return this->pushVariableReferencePartial(v, SlotRange{0, (int)v.type().slotCount()});
}
bool Generator::pushVariableReferencePartial(const VariableReference& v, SlotRange subset) {
const Variable& var = *v.variable();
SlotRange r;
if (IsUniform(var)) {
r = this->getUniformSlots(var);
SkASSERT(r.count == (int)var.type().slotCount());
r.index += subset.index;
r.count = subset.count;
fBuilder.push_uniform(r);
} else {
r = this->getVariableSlots(var);
SkASSERT(r.count == (int)var.type().slotCount());
r.index += subset.index;
r.count = subset.count;
fBuilder.push_slots(r);
}
return true;
}
bool Generator::writeProgram(const FunctionDefinition& function) {
fCurrentFunction = &function;
if (fDebugTrace) {
// Copy the program source into the debug info so that it will be written in the trace file.
fDebugTrace->setSource(*fProgram.fSource);
}
// Assign slots to the parameters of main; copy src and dst into those slots as appropriate.
for (const SkSL::Variable* param : function.declaration().parameters()) {
switch (param->modifiers().fLayout.fBuiltin) {
case SK_MAIN_COORDS_BUILTIN: {
// Coordinates are passed via RG.
SlotRange fragCoord = this->getVariableSlots(*param);
SkASSERT(fragCoord.count == 2);
fBuilder.store_src_rg(fragCoord);
break;
}
case SK_INPUT_COLOR_BUILTIN: {
// Input colors are passed via RGBA.
SlotRange srcColor = this->getVariableSlots(*param);
SkASSERT(srcColor.count == 4);
fBuilder.store_src(srcColor);
break;
}
case SK_DEST_COLOR_BUILTIN: {
// Dest colors are passed via dRGBA.
SlotRange destColor = this->getVariableSlots(*param);
SkASSERT(destColor.count == 4);
fBuilder.store_dst(destColor);
break;
}
default: {
SkDEBUGFAIL("Invalid parameter to main()");
return unsupported();
}
}
}
// Initialize the program.
fBuilder.init_lane_masks();
// Emit global variables.
if (!this->writeGlobals()) {
return unsupported();
}
// Invoke main().
if (this->needsReturnMask()) {
fBuilder.enableExecutionMaskWrites();
}
std::optional<SlotRange> mainResult = this->writeFunction(function, function);
if (!mainResult.has_value()) {
return unsupported();
}
if (this->needsReturnMask()) {
fBuilder.disableExecutionMaskWrites();
}
// Move the result of main() from slots into RGBA. Allow dRGBA to remain in a trashed state.
SkASSERT(mainResult->count == 4);
fBuilder.load_src(*mainResult);
return true;
}
std::unique_ptr<RP::Program> Generator::finish() {
return fBuilder.finish(fProgramSlots.slotCount(), fUniformSlots.slotCount(), fDebugTrace);
}
} // namespace RP
std::unique_ptr<RP::Program> MakeRasterPipelineProgram(const SkSL::Program& program,
const FunctionDefinition& function,
SkRPDebugTrace* debugTrace) {
RP::Generator generator(program, debugTrace);
if (!generator.writeProgram(function)) {
return nullptr;
}
return generator.finish();
}
} // namespace SkSL