darling-JavaScriptCore/b3/B3StackmapValue.h

322 lines
13 KiB
C++

/*
* Copyright (C) 2015-2018 Apple Inc. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY APPLE INC. ``AS IS'' AND ANY
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once
#if ENABLE(B3_JIT)
#include "B3ConstrainedValue.h"
#include "B3Value.h"
#include "B3ValueRep.h"
#include "RegisterSet.h"
#include <wtf/SharedTask.h>
namespace JSC {
class CCallHelpers;
namespace B3 {
class StackmapGenerationParams;
typedef void StackmapGeneratorFunction(CCallHelpers&, const StackmapGenerationParams&);
typedef SharedTask<StackmapGeneratorFunction> StackmapGenerator;
class JS_EXPORT_PRIVATE StackmapValue : public Value {
public:
static bool accepts(Kind kind)
{
// This needs to include opcodes of all subclasses.
switch (kind.opcode()) {
case CheckAdd:
case CheckSub:
case CheckMul:
case Check:
case Patchpoint:
return true;
default:
return false;
}
}
~StackmapValue() override;
// Use this to add children.
void append(const ConstrainedValue& value)
{
ASSERT(value.value()->type().isNumeric());
append(value.value(), value.rep());
}
void append(Value*, const ValueRep&);
template<typename VectorType>
void appendVector(const VectorType& vector)
{
for (const auto& value : vector)
append(value);
}
// Helper for appending a bunch of values with some ValueRep.
template<typename VectorType>
void appendVectorWithRep(const VectorType& vector, const ValueRep& rep)
{
for (Value* value : vector)
append(value, rep);
}
// Helper for appending cold any's. This often used by clients to implement OSR.
template<typename VectorType>
void appendColdAnys(const VectorType& vector)
{
appendVectorWithRep(vector, ValueRep::ColdAny);
}
template<typename VectorType>
void appendLateColdAnys(const VectorType& vector)
{
appendVectorWithRep(vector, ValueRep::LateColdAny);
}
// This is a helper for something you might do a lot of: append a value that should be constrained
// to SomeRegister.
void appendSomeRegister(Value*);
void appendSomeRegisterWithClobber(Value*);
const Vector<ValueRep>& reps() const { return m_reps; }
// Stackmaps allow you to specify that the operation may clobber some registers. Clobbering a register
// means that the operation appears to store a value into the register, but the compiler doesn't
// assume to know anything about what kind of value might have been stored. In B3's model of
// execution, registers are read or written at instruction boundaries rather than inside the
// instructions themselves. A register could be read or written immediately before the instruction
// executes, or immediately after. Note that at a boundary between instruction A and instruction B we
// simultaneously look at what A does after it executes and what B does before it executes. This is
// because when the compiler considers what happens to registers, it views the boundary between two
// instructions as a kind of atomic point where the late effects of A happen at the same time as the
// early effects of B.
//
// The compiler views a stackmap as a single instruction, even though of course the stackmap may be
// composed of any number of instructions (if it's a Patchpoint). You can claim that a stackmap value
// clobbers a set of registers before the stackmap's instruction or after. Clobbering before is called
// early clobber, while clobbering after is called late clobber.
//
// This is quite flexible but it has its limitations. Any register listed as an early clobber will
// interfere with all uses of the stackmap. Any register listed as a late clobber will interfere with
// all defs of the stackmap (i.e. the result). This means that it's currently not possible to claim
// to clobber a register while still allowing that register to be used for both an input and an output
// of the instruction. It just so happens that B3's sole client (the FTL) currently never wants to
// convey such a constraint, but it will want it eventually (FIXME:
// https://bugs.webkit.org/show_bug.cgi?id=151823).
//
// Note that a common use case of early clobber sets is to indicate that this is the set of registers
// that shall not be used for inputs to the value. But B3 supports two different ways of specifying
// this, the other being LateUse in combination with late clobber (not yet available to stackmaps
// directly, FIXME: https://bugs.webkit.org/show_bug.cgi?id=151335). A late use makes the use of that
// value appear to happen after the instruction. This means that a late use cannot use the same
// register as the result and it cannot use the same register as either early or late clobbered
// registers. Late uses are usually a better way of saying that a clobbered register cannot be used
// for an input. Early clobber means that some register(s) interfere with *all* inputs, while LateUse
// means that some value interferes with whatever is live after the instruction. Below is a list of
// examples of how the FTL can handle its various kinds of scenarios using a combination of early
// clobber, late clobber, and late use. These examples are for X86_64, w.l.o.g.
//
// Basic ById patchpoint: Early and late clobber of r11. Early clobber prevents any inputs from using
// r11 since that would mess with the MacroAssembler's assumptions when we
// AllowMacroScratchRegisterUsage. Late clobber tells B3 that the patchpoint may overwrite r11.
//
// ById patchpoint in a try block with some live state: This might throw an exception after already
// assigning to the result. So, this should LateUse all stackmap values to ensure that the stackmap
// values don't interfere with the result. Note that we do not LateUse the non-OSR inputs of the ById
// since LateUse implies that the use is cold: the register allocator will assume that the use is not
// important for the critical path. Also, early and late clobber of r11.
//
// Basic ByIdFlush patchpoint: We could do Flush the same way we did it with LLVM: ignore it and let
// PolymorphicAccess figure it out. Or, we could add internal clobber support (FIXME:
// https://bugs.webkit.org/show_bug.cgi?id=151823). Or, we could do it by early clobbering r11, late
// clobbering all volatile registers, and constraining the result to some register. Or, we could do
// that but leave the result constrained to SomeRegister, which will cause it to use a callee-save
// register. Internal clobber support would allow us to use SomeRegister while getting the result into
// a volatile register.
//
// ByIdFlush patchpoint in a try block with some live state: LateUse all for-OSR stackmap values,
// early clobber of r11 to prevent the other inputs from using r11, and late clobber of all volatile
// registers to make way for the call. To handle the result, we could do any of what is listed in the
// previous paragraph.
//
// Basic JS call: Force all non-OSR inputs into specific locations (register, stack, whatever).
// All volatile registers are late-clobbered. The output is constrained to a register as well.
//
// JS call in a try block with some live state: LateUse all for-OSR stackmap values, fully constrain
// all non-OSR inputs and the result, and late clobber all volatile registers.
//
// JS tail call: Pass all inputs as a warm variant of Any (FIXME:
// https://bugs.webkit.org/show_bug.cgi?id=151811).
//
// Note that we cannot yet do all of these things because although Air already supports all of these
// various forms of uses (LateUse and warm unconstrained use), B3 doesn't yet expose all of it. The
// bugs are:
// https://bugs.webkit.org/show_bug.cgi?id=151335 (LateUse)
// https://bugs.webkit.org/show_bug.cgi?id=151811 (warm Any)
void clobberEarly(const RegisterSet& set)
{
m_earlyClobbered.merge(set);
}
void clobberLate(const RegisterSet& set)
{
m_lateClobbered.merge(set);
}
void clobber(const RegisterSet& set)
{
clobberEarly(set);
clobberLate(set);
}
RegisterSet& earlyClobbered() { return m_earlyClobbered; }
RegisterSet& lateClobbered() { return m_lateClobbered; }
const RegisterSet& earlyClobbered() const { return m_earlyClobbered; }
const RegisterSet& lateClobbered() const { return m_lateClobbered; }
void setGenerator(RefPtr<StackmapGenerator> generator)
{
m_generator = generator;
}
template<typename Functor>
void setGenerator(const Functor& functor)
{
m_generator = createSharedTask<StackmapGeneratorFunction>(functor);
}
RefPtr<StackmapGenerator> generator() const { return m_generator; }
ConstrainedValue constrainedChild(unsigned index) const
{
return ConstrainedValue(child(index), index < m_reps.size() ? m_reps[index] : ValueRep::ColdAny);
}
void setConstrainedChild(unsigned index, const ConstrainedValue&);
void setConstraint(unsigned index, const ValueRep&);
class ConstrainedValueCollection {
public:
ConstrainedValueCollection(const StackmapValue& value)
: m_value(value)
{
}
unsigned size() const { return m_value.numChildren(); }
ConstrainedValue at(unsigned index) const { return m_value.constrainedChild(index); }
ConstrainedValue operator[](unsigned index) const { return at(index); }
class iterator {
public:
using iterator_category = std::forward_iterator_tag;
using value_type = ConstrainedValue;
using difference_type = int;
using pointer = void;
using reference = ConstrainedValue;
iterator()
: m_collection(nullptr)
, m_index(0)
{
}
iterator(const ConstrainedValueCollection& collection, unsigned index)
: m_collection(&collection)
, m_index(index)
{
}
ConstrainedValue operator*() const
{
return m_collection->at(m_index);
}
iterator& operator++()
{
m_index++;
return *this;
}
bool operator==(const iterator& other) const
{
ASSERT(m_collection == other.m_collection);
return m_index == other.m_index;
}
bool operator!=(const iterator& other) const
{
return !(*this == other);
}
private:
const ConstrainedValueCollection* m_collection;
unsigned m_index;
};
iterator begin() const { return iterator(*this, 0); }
iterator end() const { return iterator(*this, size()); }
private:
const StackmapValue& m_value;
};
ConstrainedValueCollection constrainedChildren() const
{
return ConstrainedValueCollection(*this);
}
B3_SPECIALIZE_VALUE_FOR_VARARGS_CHILDREN
protected:
void dumpChildren(CommaPrinter&, PrintStream&) const override;
void dumpMeta(CommaPrinter&, PrintStream&) const override;
StackmapValue(CheckedOpcodeTag, Kind, Type, Origin);
private:
friend class CheckSpecial;
friend class PatchpointSpecial;
friend class StackmapGenerationParams;
friend class StackmapSpecial;
Vector<ValueRep> m_reps;
RefPtr<StackmapGenerator> m_generator;
RegisterSet m_earlyClobbered;
RegisterSet m_lateClobbered;
RegisterSet m_usedRegisters; // Stackmaps could be further duplicated by Air, but that's unlikely, so we just merge the used registers sets if that were to happen.
};
} } // namespace JSC::B3
#endif // ENABLE(B3_JIT)