B3 opcodes should leave room for flags
https://bugs.webkit.org/show_bug.cgi?id=162692
Reviewed by Keith Miller.
Source/JavaScriptCore:
It used to be that the main thing that determined what a Value did was the opcode. The
Opcode was how you knew what subclass of Value you had. The opcode told you what the Value
actually did. This change replaces Opcode with Kind, which is a tuple of opcode and other
stuff.
Opcodes are great, and that's how most compilers work. But opcodes are one-dimensional. Here
is how this manifests. Say you have an opcode, like Load. You will be happy if your IR has
one Load opcode. But then, you might add Load8S/Load8Z/Load16S/Load16Z opcodes, as we have
done in B3. B3 has one dimension of Load opcodes, which determines something like the C type
of the load. But in the very near future, we will want to add two more dimensions to Loads:
- A flag to say if the load traps.
- A flag to say if the load has acquire semantics.
Mapping these three dimensions (type, trap, acquire) onto the one-dimensional Opcode space
would create mayham: Load8S, Load8STrap, Load8SAcquire, Load8STrapAcquire, Load8Z,
Load8ZTrap, etc.
This happens in other parts of the IR. For example, we have a dimension of arithmetic
operations: add, sub, mul, div, mod, etc. Then we have the chill flag. But since opcodes
are one-dimensional, that means having ChillDiv and ChillMod, and tons of places in the
compiler that case on both Div and ChillDiv, or case on both Mod and ChillMod, since they
are only interested in the kind of arithmetic being done and not the chillness.
Though the examples all involve bits (chill or not, trapping or not, etc), I can imagine
other properties that behave more like small enums, like if we fill out more memory ordering
modes other than just "acquire? yes/no". There will eventually have to be something like a
std::memory_order associated with memory accesses.
One approach to this problem is to have a Value subclass that contains fields with the meta
data. I don't like this for two reasons:
- In bug 162688, I want to make trapping memory accesses have stackmaps. This means that a
trapping memory access would have a different Value subclass than a non-trapping memory
access. So, this meta-data needs to channel into ValueType::accepts(). Currently that
takes Opcode and nothing else.
- Compiler IRs are all about making common tasks easy. If it becomes commonplace for opcodes
to require a custom Value subclass just for a bit then that's not very easy.
This change addresses this problem by making the compiler pass around Kinds rather than
Opcodes. A Kind contains an Opcode as well as any number of opcode-specific bits. This
change demonstrates how Kind should be used by converting chillness to it. Kind has
hasIsChill(), isChill(), and setIsChill() methods. hasIsChill() is true only for Div and
Mod. setIsChill() asserts if !hasIsChill(). If you want to create a Chill Div, you say
chill(Div). IR dumps will print it like this:
Int32 @38 = Div<Chill>(@36, @37, DFG:@24, ControlDependent)
Where "Div<Chill>" is how a Kind that hasExtraBits() dumps itself. If a Kind does not
hasExtraBits() (the normal case) then it dumps like a normal Opcode (without the "<>").
I replaced many uses of Opcode with Kind. New code has to be mindful that Opcode may not be
the right way to summarize what a value does, and so in many cases it's better to carry
around a Kind instead - especially if you will use it to stamp out new Values. Opcode is no
longer sufficient to perform a dynamic Value cast, since that code now uses Kind. ValueKey
now wants a Kind instead of an Opcode. All Value constructors now take Kind instead of
Opcode. But most opcodes don't get any extra Kind bits, and so the code that operates on
those opcodes is largely unchanged.
- CMakeLists.txt:
- JavaScriptCore.xcodeproj/project.pbxproj:
- b3/B3ArgumentRegValue.h:
- b3/B3CCallValue.h:
- b3/B3CheckValue.cpp:
(JSC::B3::CheckValue::convertToAdd):
(JSC::B3::CheckValue::CheckValue):
(JSC::B3::CheckValue::accepts):
- b3/B3Const32Value.h:
- b3/B3Const64Value.h:
- b3/B3ConstDoubleValue.h:
- b3/B3ConstFloatValue.h:
- b3/B3FenceValue.h:
- b3/B3Kind.cpp: Added.
(JSC::B3::Kind::dump):
(JSC::B3::Kind::Kind):
(JSC::B3::Kind::opcode):
(JSC::B3::Kind::setOpcode):
(JSC::B3::Kind::hasExtraBits):
(JSC::B3::Kind::hasIsChill):
(JSC::B3::Kind::isChill):
(JSC::B3::Kind::setIsChill):
(JSC::B3::Kind::operator==):
(JSC::B3::Kind::operator!=):
(JSC::B3::Kind::hash):
(JSC::B3::Kind::isHashTableDeletedValue):
(JSC::B3::chill):
(JSC::B3::KindHash::hash):
(JSC::B3::KindHash::equal):
- b3/B3LowerMacros.cpp:
- b3/B3LowerToAir.cpp:
(JSC::B3::Air::LowerToAir::lower):
- b3/B3MemoryValue.h:
- b3/B3Opcode.cpp:
(WTF::printInternal):
- b3/B3Opcode.h:
- b3/B3PatchpointValue.h:
(JSC::B3::PatchpointValue::accepts):
- b3/B3ReduceStrength.cpp:
- b3/B3SlotBaseValue.h:
- b3/B3StackmapValue.cpp:
(JSC::B3::StackmapValue::StackmapValue):
- b3/B3StackmapValue.h:
- b3/B3SwitchValue.h:
(JSC::B3::SwitchValue::accepts):
- b3/B3UpsilonValue.h:
- b3/B3Validate.cpp:
- b3/B3Value.cpp:
(JSC::B3::Value::dump):
(JSC::B3::Value::deepDump):
(JSC::B3::Value::invertedCompare):
(JSC::B3::Value::effects):
(JSC::B3::Value::key):
(JSC::B3::Value::typeFor):
(JSC::B3::Value::badKind):
(JSC::B3::Value::badOpcode): Deleted.
- b3/B3Value.h:
- b3/B3ValueInlines.h:
(JSC::B3::Value::as):
(JSC::B3::ValueKey::dump):
(JSC::B3::ValueKey::materialize):
(JSC::B3::ValueKey::ValueKey):
(JSC::B3::ValueKey::kind):
(JSC::B3::ValueKey::opcode):
(JSC::B3::ValueKey::operator==):
(JSC::B3::ValueKey::hash):
(JSC::B3::ValueKey::ValueKey):
(JSC::B3::VariableValue::VariableValue):
- b3/B3VariableValue.h:
- b3/testb3.cpp:
(JSC::B3::testChillDiv):
(JSC::B3::testChillDivTwice):
(JSC::B3::testChillDiv64):
(JSC::B3::testChillModArg):
(JSC::B3::testChillModArgs):
(JSC::B3::testChillModImms):
(JSC::B3::testChillModArg32):
(JSC::B3::testChillModArgs32):
(JSC::B3::testChillModImms32):
(JSC::B3::testSwitchChillDiv):
(JSC::B3::testEntrySwitchWithCommonPaths):
(JSC::B3::testEntrySwitchWithCommonPathsAndNonTrivialEntrypoint):
(JSC::FTL::Output::chillDiv):
(JSC::FTL::Output::chillMod):
Websites/webkit.org:
Updated the documentation to talk about Kind and the isChill bit, and to remove
ChillDiv/ChillMod.
- docs/b3/intermediate-representation.html:
