Bare Bones Backend / B3 Intermediate Representation

B3 IR is a C-like SSA representation of a procedure. A procedure has a root block at which it starts execution when it is invoked. A procedure does not have to terminate, but if it does, then it can be either due to a Return, which gracefully returns some value, or by a side-exit at designated instructions. B3 gives the client a lot of flexibility to implement many different kinds of side-exits.

B3 is designed to represent procedures for the purpose of transforming them. Knowing what transformations are legal requires knowing what a procedure does. A transformation is valid if it does not change the observable behavior of a procedure. This document tells you what B3 procedures do by telling you what each construct in B3 IR does.

Procedure

The parent object of all things in B3 is the Procedure. Every time you want to compile something, you start by creating a Procedure. The lifecycle of a Procedure is usually:

Create the Procedure.
Populate the Procedure with code.
Use either the high-level Compilation API or the low-level generation API.

The act of compiling the Procedure changes it in-place, making it unsuitable for compiling again. Always create a new Procedure every time you want to compile something.

Types

B3 has a trivial type system with only five types:

Void: Used to say that an instruction does not return a value.
Int32: 32-bit integer. Integers don't have sign, but operations on them do.
Int64: 64-bit integer.
Float: 32-bit binary floating point number.
Double: 64-bit binary floating point number.

B3 does not have a pointer type. Instead, the B3::pointerType() function will return either Int32 or Int64 depending on which kind of integer can be used to represent a pointer on the current platform. It's not a goal of B3 to support hardware targets that require pointers and integers to be segregated. It's not a goal of B3 to support GC (garbage collection) roots as a separate type, since JSC uses Bartlett-style conservative root scanning. This doesn't preclude any mainstream garbage collection algorithms, including copying, generational, or concurrent collectors, and frees up the compiler to perform more optimizations.

Values

Variables, and the instructions that define them, are represented using the Value object. The Value object has a return type, a kind, and zero or more children. Children are references to other Values. Those values are used as input to the instruction that computes this value.

The value kind is a combination of an opcode and optional flags. The flags allow a single opcode to have many variants. For example, Div and Mod may have the Chill flag set to indicate that they should not trap on corner cases. Alternatively, Load/Store opcodes may have the Traps flag set to indicate that they should trap deterministically.

Values also have a unique 32-bit index that is used as the name.

Example:

Int32 @3 = Add(@1, @2)

This represents a single Value instance. Its index is 3. It is an Int32. The opcode is Add, and its children are @1 and @2.

Values may also have additional meta-data. We use special subclasses of the B3::Value class for values that need meta-data. For example, the Load value needs a 32-bit offset for the load. We use the MemoryValue class for memory-accessing values, which all have such an offset.

Stack Slots

B3 exposes the concept of stack-allocated data and gives the client a lot of control. By default, stack slots get allocated wherever B3 chooses. It will try to pack them as much as possible. After compilation is done, you can retrieve each stack slot's location in the form of an offset from the frame pointer.

You can force stack slots to end up at a particular offset from the frame pointer, though this is very dangerous. For example, B3 assumes that it can use the slots closest to the frame pointer for callee-saves, and currently when you force something to a particular frame pointer offset, there is no mechanism to notice that this space is also being used for callee-saves. Therefore, we recommend not using the frame pointer offset forcing feature unless you know a lot about the ABI and you have no other choice.

Variables

Sometimes, SSA is inconvenient. For example, it's hard to do path specialization over SSA. B3 has the concept of Variables as a fall-back. The backend knows how to handle them and will coalesce and copy-propagate them. Inside the B3 optimizer, there is a classic SSA builder that eliminates variables and builds SSA in their place.

You can create Variables by using Procedure::addVariable(), and then you can access them using the Get and Set opcodes.

The fixSSA() phase will convert variables to SSA. If you use a lot of variables in your input to B3, it's a good idea to run fixSSA() manually before running the compiler. The default optimizer only runs fixSSA() towards the middle of optimizations. Passing non-SSA code as input to the optimizer may render the early phases ineffective. Fortunately, B3 phases are super easy to run. The following runs SSA fix-up on a Procedure named "proc":

fixSSA(proc);

Control flow

B3 represents control flow using basic blocks. Each basic block may have zero or more predecessors. Each basic block may have zero or more successors. The meaning of the successors is determined by the basic block's last Value, which must be a terminal. A value is terminal if:

value->effects().terminal

Some opcodes are definitely terminal, like Jump, Branch, Oops, Return, and Switch. But a value with the Patchpoint opcode may or may not be terminal. In general it's necessary to check the terminal bit as shown above.

Each basic block contains a Vector<Value*> as the contents of the block. Control flow inside the block is implicit based on the order of Values in the vector.

Memory model

B3 uses a secondary address space of abstract heaps to reason about aliasing and concurrency. This address space is 32-bit and we point at it using the HeapRange type. If you never supply HeapRanges to your memory operations (the default), they will have [0, UINT_MAX] as their range, indicating that the operation may conservatively write to all 2³² abstract heaps. Two memory operations are said to alias if they access the same abstract heaps. A simple example of using HeapRanges is with loads and stores, but you can also supply HeapRanges as bounds on the effects of calls and patchpoints.

Fences are modeled as phantom effects that we call fence ranges. For example, a load fence could be represented as a patchpoint that writes to [0, UINT_MAX]. B3 models load fences, store fences, store-load fences, acquire fences, and release fences using fence ranges. The absence of a fence range means that there is no fence.

Acquire fences are modeled as phantom writes to memory, while release fences are modeled as phantom reads from memory. By itself this does not permit B3 to do all of the reorderings that acquire/release allow. B3 is therefore allowed to treat loads/stores with fence ranges more precisely. The phantom write effects of a fenced load cannot write to anything that is read or written after the fence. The phantom read effects of a fenced store cannot read anything that is written after the fence.

Applying a fence range to a load or store only makes that access a half-fence: load-with-fence is an acquire fence that only prevents later operations from being hoisted and store-with-fence is a release fence that only prevents earlier operations from being sunk. The canonical way to express a fully fenced load (which is fenced in both directions) is to use AtomicXchgAdd(0, ptr) with a non-empty fence range. The canonical way to express a fully fenced store is to use AtomicXchg(value, ptr) with a non-empty fence range.

Opcodes

This section describes opcodes in the following format:

Int32 Foo(Int64, Double): This describes an opcode named Foo that uses Int32 as its return type and takes two children - one of type Int64 and another of type Double.

We sometimes use the wildcard type T to represent polymorphic operations, like "T Add(T, T)". This means that the value must take two children of the same type and returns a value of that type. We use the type IntPtr to mean either Int32, or Int64, depending on the platform.

Some opcodes can have some flags set. A description of flags follows the description of opcodes.

Opcode descriptions

Void Nop()

The empty value. Instead of removing Values from basic blocks, most optimizations convert them to Nops. Various phases run fix-up where all Nops are removed in one pass. It's common to see Nops in intermediate versions of B3 IR during optimizations. Nops never lead to any code being generated and they do not impede optimizations, so they are usually harmless. You can convert a Value to a Nop by doing convertToNop().

T Identity(T)

Returns the passed value. May be used for any type except Void. Instead of replacing all uses of a Value with a different Value, most optimizations convert them to Identity. Various phases run fix-up where all uses of Identity are replaced with the Identity's child (transitively, so Identity(Identity(Identity(@x))) is changed to just @x). Even the instruction selector "sees through" Identity. You can remove all references to Identity in any value by calling Value::performSubstitution(). You can convert a Value to an Identity by doing convertToIdentity(otherValue). If the value is Void, convertToIdentity() converts it to a Nop instead.

T Opaque(T)

Returns the passed value. May be used for any type except Void. This opcode is provided as a hack for avoiding B3 optimizations that the client finds unprofitable. B3 treats Opaque as an identity only during instruction selection. All prior optimizations (including all B3 IR optimizations) do not know what Opaque does, other than Opaque(x) == Opaque(x) and Opaque(Opaque(x)) == Opaque(x).

Int32 Const32(constant)

32-bit integer constant. Must use the Const32Value class, which has space for the int32_t constant.

Int64 Const64(constant)

64-bit integer constant. Must use the Const64Value class, which has space for the int64_t constant.

Float ConstFloat(constant)

Float constant. Must use the ConstFloatValue class, which has space for the float constant.

Double ConstDouble(constant)

Double constant. Must use the ConstDoubleValue class, which has space for the double constant.

Void Set(value, variable)

Assigns the given value to the given Variable. Must use the VariableValue class.

T Get(variable)

Returns the current value of the given Variable. Its return type depends on the variable. Must use the VariableValue class.

IntPtr SlotBase(stackSlot)

Returns a pointer to the base of the given stack slot. Must use the SlotBaseValue class.

IntPtr|Double ArgumentReg(%register)

Returns the value that the given register had at the prologue of the procedure. It returns IntPtr for general-purpose registers and Double for FPRs. Must use the ArgumentRegValue class.

IntPtr FramePointer()

Returns the value of the frame pointer register. B3 procedures alway use a frame pointer ABI, and the frame pointer is always guaranteed to have this value anywhere inside the procedure.

T Add(T, T)

Works with any type except Void. For integer types, this represents addition with wrap-around semantics. For floating point types, this represents addition according to the IEEE 854 spec. B3 does not have any notion of "fast math". A transformation over floating point code is valid if the new code produces exactly the same value, bit for bit.

T Sub(T, T)

Works with any type except Void. For integer types, this represents subtraction with wrap-around semantics. For floating point types, this represents subtraction according to the IEEE 854 spec.

T Mul(T, T)

Works with any type except Void. For integer types, this represents multiplication with wrap-around semantics. For floating point types, this represents multiplication according to the IEEE 854 spec.

T Div(T, T)

Works with any type except Void. For integer types, this represents signed division with round-to-zero. By default, its behavior is undefined for x/0 or -2³¹/-1. For floating point types, this represents division according to the IEEE 854 spec. Integer Div may have the Chill flag set.

T Mod(T, T)

Works with any type except Void. For integer types, this represents signed modulo. By default, its behavior is undefined for x%0 or -2³¹%-1. For floating point types, this represents modulo according to "fmod()". Integer Mod may have the Chill flag set.

T Neg(T)

Works with any type except Void. For integer types, this represents twos-complement negation. For floating point types, this represents negation according to the IEEE spec.

T BitAnd(T, T)

Bitwise and. Valid for any type except Void.

T BitOr(T, T)

Bitwise or. Valid for any type except Void.

T BitXor(T, T)

Bitwise xor. Valid for any type except Void.

T Shl(T, Int32)

Shift left. Valid for Int32 and Int64. The shift amount is always Int32. Only the low 31 bits of the shift amount are used for Int32. Only the low 63 bits of the shift amount are used for Int64.

T SShr(T, Int32)

Shift right with sign extension. Valid for Int32 and Int64. The shift amount is always Int32. Only the low 31 bits of the shift amount are used for Int32. Only the low 63 bits of the shift amount are used for Int64.

T ZShr(T, Int32)

Shift right with zero extension. Valid for Int32 and Int64. The shift amount is always Int32. Only the low 31 bits of the shift amount are used for Int32. Only the low 63 bits of the shift amount are used for Int64.

T RotR(T, Int32)

Rotate right. Valid for Int32 and Int64. The shift amount is always Int32. Only the low 31 bits of the shift amount are used for Int32. Only the low 63 bits of the shift amount are used for Int64.

T RotL(T, Int32)

Rotate left. Valid for Int32 and Int64. The shift amount is always Int32. Only the low 31 bits of the shift amount are used for Int32. Only the low 63 bits of the shift amount are used for Int64.

T Clz(T)

Count leading zeroes. Valid for Int32 and Int64.

T Abs(T)

Absolute value. Valid for Float and Double.

T Ceil(T)

Ceiling. Valid for Float and Double.

T Floor(T)

Flooring. Valid for Float and Double.

T Sqrt(T)

Square root. Valid for Float and Double.

U BitwiseCast(T)

If T is Int32 or Int64, it returns the bitwise-corresponding Float or Double, respectively. If T is Float or Double, it returns the bitwise-corresponding Int32 or Int64, respectively.

Int32 SExt8(Int32)

Fills the top 24 bits of the integer with the low byte's sign extension.

Int32 SExt16(Int32)

Fills the top 16 bits of the integer with the low short's sign extension.

Int64 SExt32(Int32)

Returns a 64-bit integer such that the low 32 bits are the given Int32 value and the high 32 bits are its sign extension.

Int64 ZExt32(Int32)

Returns a 64-bit integer such that the low 32 bits are the given Int32 value and the high 32 bits are zero.

U Trunc(T)

Returns the low 32 bits of the 64-bit value. If T is Int64 then U is Int32. If T is Double then U is Float.

Double IToD(T)

Converts the given integer into a double. Value for Int32 or Int64 inputs.

Double FloatToDouble(Float)

Converts the given float into a double.

Float DoubleToFloat(Double)

Converts the given double into a float.

Int32 Equal(T, T)

Compares the two values. If they are equal, return 1; else return 0. Valid for all types except Void. Integer comparisons simply compare all bits. Floating point comparisons mostly compare bits, but have some corner cases: positive zero and negative zero are considered equal, and they return false when either value is NaN.

Int32 NotEqual(T, T)

The opposite of Equal(). NotEqual(@x, @y) yields the same result as BitXor(Equal(@x, @y), 1).

Int32 LessThan(T, T)

Returns 1 if the left value is less than the right one, 0 otherwise. Does a signed comparison for integers. For floating point comparisons, this has the usual caveats with respect to negative zero and NaN.

Int32 GreaterThan(T, T)

Returns 1 if the left value is greater than the right one, 0 otherwise. Does a signed comparison for integers. For floating point comparisons, this has the usual caveats with respect to negative zero and NaN.

Int32 LessEqual(T, T)

Returns 1 if the left value is less than or equal to the right one, 0 otherwise. Does a signed comparison for integers. For floating point comparisons, this has the usual caveats with respect to negative zero and NaN.

Int32 GreaterEqual(T, T)

Returns 1 if the left value is greater than or equal to the right one, 0 otherwise. Does a signed comparison for integers. For floating point comparisons, this has the usual caveats with respect to negative zero and NaN.

Int32 Above(T, T)

Unsigned integer comparison, valid for Int32 and Int64 only. Returns 1 if the left value is unsigned-greater-than the right one, 0 otherwise.

Int32 Below(T, T)

Unsigned integer comparison, valid for Int32 and Int64 only. Returns 1 if the left value is unsigned-less-than the right one, 0 otherwise.

Int32 AboveEqual(T, T)

Unsigned integer comparison, valid for Int32 and Int64 only. Returns 1 if the left value is unsigned-greater-than-or-equal the right one, 0 otherwise.

Int32 BelowEqual(T, T)

Unsigned integer comparison, valid for Int32 and Int64 only. Returns 1 if the left value is unsigned-less-than-or-equal the right one, 0 otherwise.

Int32 EqualOrUnordered(T, T)

Floating point comparison, valid for Float and Double only. Returns 1 if the left value is equal to the right one or if either value is NaN. Returns 0 otherwise.

T Select(U, T, T)

Returns either the second child or the third child. T can be any type except Void. U can be either Int32 or Int64. If the first child is non-zero, returns the second child. Otherwise returns the third child.

Int32 Load8Z(IntPtr, offset)

Loads a byte from the address, which is computed by adding the compile-time 32-bit signed integer offset to the child value. Zero extends the loaded byte, so the high 24 bits are all zero. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a load acquire.

Int32 Load8S(IntPtr, offset)

Loads a byte from the address, which is computed by adding the compile-time 32-bit signed integer offset to the child value. Sign extends the loaded byte. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a load acquire.

Int32 Load16Z(IntPtr, offset)

Loads a 16-bit integer from the address, which is computed by adding the compile-time 32-bit signed integer offset to the child value. Zero extends the loaded 16-bit integer, so the high 16 bits are all zero. Misaligned loads are not penalized. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a load acquire.

Int32 Load16S(IntPtr, offset)

Loads a 16-bit integer from the address, which is computed by adding the compile-time 32-bit signed integer offset to the child value. Sign extends the loaded 16-bit integer. Misaligned loads are not penalized. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a load acquire.

T Load(IntPtr, offset)

Valid for any type except Void. Loads a value of that type from the address, which is computed by adding the compile-time 32-bit signed integer offset to the child value. Misaligned loads are not penalized. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a load acquire.

Void Store8(Int32, IntPtr, offset)

Stores a the low byte of the first child into the address computed by adding the compile-time 32-bit signed integer offset to the second child. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a store release.

Void Store16(Int32, IntPtr, offset)

Stores a the low 16 bits of the first child into the address computed by adding the compile-time 32-bit signed integer offset to the second child. Misaligned stores are not penalized. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a store release.

Void Store(T, IntPtr, offset)

Stores the value in the first child into the address computed by adding the compile-time 32-bit signed integer offset to the second child. Misaligned stores are not penalized. Must use the MemoryValue class. May have the Traps flag set. May set a fence range to turn this into a store release.

Int32 AtomicWeakCAS(T expectedValue, T newValue, IntPtr ptr)

Performs a weak CAS (compare-and-swap). Returns a boolean (0 or 1) to indicate the result (failure or success, respectively). May spuriously fail, in which case it will do nothing and return 0. You can supply a fence range to indicate that the CAS is fenced. Fenced CAS has both acquire and release fences, and claims to both read and write the full fence range in addition to whatever primary range is supplied for the CAS itself. AtomicWeakCAS only has fencing behavior in the case that it succeeds. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations.

T AtomicStrongCAS(T expectedValue, T newValue, IntPtr ptr)

Performs a strong CAS (compare-and-swap). Returns the old value before the CAS. The instruction selector is smart about Equal(AtomicStrongCAS(expected, ...), expected) patterns, so this opcode is also the best way to do a strong CAS that returns a boolean - just use Equal to compare the result to expected. You can supply a fence range to indicate that the CAS is fenced. Fenced CAS has both acquire and release fences, and claims to both read and write the full fence range in addition to whatever primary range is supplied for the CAS itself. AtomicStrongCAS has the same fencing on failure and success. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchgAdd(T, IntPtr)

Atomically adds a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchgAnd(T, IntPtr)

Atomically bitwise-ands a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchgOr(T, IntPtr)

Atomically bitwise-ors a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchgSub(T, IntPtr)

Atomically subtracts a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchgXor(T, IntPtr)

Atomically bitwise-xors a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T AtomicXchg(T, IntPtr)

Atomically stores a value to the memory location and returns the old value. Is allowed to have a fence range, which causes it to have acquire/release fencing. Atomic operations take a width parameter, allowing them to operate over 8-bit, 16-bit, 32-bit, or 64-bit integer memory locations. Returns a sign-extended result for 8-bit or 16-bit widths.

T Depend(T)

Only valid for integer types. This is guaranteed to return zero by xoring the input with itself, and the B3 compiler guarantees that it will not leverage this knowledge for any optimization. On x86, this is lowered to a Fence with read=Bottom, write=Top (i.e. a load-load fence) and then it's constant-folded to zero. On ARM, this is preserved through the compiler pipeline all the way to the MacroAssembler, which turns this into a eor. Using Depend is the most efficient way to create load-load dependencies, regarldess of platform. It allows B3's CSE to work for surrounding loads and it even supports CSEing the dependency itself - so two identical load chains with no interleaving effects can be turned into one. On ARM, it also avoids emitting an expensive dmb ish load-load fence. Because of the benefits for CSE, it's recommended to use Depend for load-load dependencies even if you're only targeting x86.

IntPtr WasmAddress(IntPtr, pinnedGPR)

This is used to compute the address of a wasm memory load from a pinned base GPR. Must use the WasmAddressValue class.

Void Fence()

Abstracts standalone data fences on x86 and ARM. Must use the FenceValue class, which has two additional members that configure the precise meaning of the fence: HeapRange FenceValue::read and HeapRange FenceValue::write. If the write range is empty, this is taken to be a store-store fence, which leads to no code generation on x86 and the weaker dmb ishst fence on ARM. If the write range is non-empty, this produces mfence on x86 and dmb ish on ARM. Within B3 IR, the Fence also reports the read/write in its effects. This allows you to scope the fence for the purpose of B3's load elimination. For example, you may use a Fence to protect a store from being sunk below a specific load. In that case, you can claim to read just that store's range and write that load's range.

T1 CCall(IntPtr, [T2, [T3, ...]])

Performs a C function call to the function pointed to by the first child. The types that the function takes and the type that it returns are determined by the types of the children and the type of the CCallValue. Only the first child is mandatory. Must use the CCallValue class.

T1 Patchpoint([T2, [T3, ...]])

A Patchpoint is a customizable value. Patchpoints take zero or more values of any type and return any type. A Patchpoint's behavior is determined by the generator object. The generator is a C++ lambda that gets called during code generation. It gets passed an assembler instance (specifically, CCallHelpers&) and an object describing where to find all of the input values and where to put the result. Here's an example:

PatchpointValue* patchpoint = block->appendNew<PatchpointValue>(proc, Int32, Origin());
patchpoint->append(ConstrainedValue(arg1, ValueRep::SomeRegister));
patchpoint->append(ConstrainedValue(arg2, ValueRep::SomeRegister));
patchpoint->setGenerator(
    [&] (CCallHelpers& jit, const StackmapGenerationParams& params) {
        jit.add32(params[1].gpr(), params[2].gpr(), params[0].gpr());
    });

This creates a patchpoint that just adds two numbers. The patchpoint is set to return Int32. The two child values, arg1 and arg2, are passed to the patchpoint with the SomeRegister constraint, which just requests that they get put in appropriate registers (GPR for integer values, FPR for floating point values). The generator uses the params object to figure out which registers the inputs are in (params[1] and params[2]) and which register to put the result in (params[0]). Many sophisticated constraints are possible. You can request that a child gets placed in a specific register. You can list additional registers that are clobbered - either at the top of the patchpoint (i.e. early) so that the clobbered registers interfere with the inputs, or at the bottom of the patchpoint (i.e. late) so that the clobbered registers interfere with the output. Patchpoint constraints also allow you to force values onto the call argument area of the stack. Patchpoints are powerful enough to implement custom calling conventions, inline caches, and side-exits.

A patchpoint is allowed to "side exit", i.e. abruptly exit from the procedure. If it wants to do so by returning, it can use Procedure's API for getting the callee-save register layout, unwinding the callee-saves, and then returning. More likely, the patchpoint will take some exit state as part of its arguments, and it will manipulate the call frame in place to make it look like another execution engine's call frame. This is called OSR, and JavaScriptCore does it a lot.

A patchpoint can be used as a terminal. Simply set the terminal bit:

patchpoint->effects.terminal = true;

The generator can determine where to branch by using the StackmapGenerationParams to get the set of labels for all successors. You're guaranteed that the number of successors of the patchpoint's basic block will be the same as when you created it. However, like any value, a patchpoint may be cloned. This means, for example, that if you use this to implement a table jump then the jump table must be created inside the generator, so that you get one jump table per clone.

Must use the PatchpointValue class with the Patchpoint opcode.

T CheckAdd(T, T, [T2, [T3, ...]])

Checked integer addition. Works for T = Int32 or T = Int64. First first two children are mandatory. Additional children are optional. All of the Check instructions take a generator and value constraints like a Patchpoint. In the case of a CheckAdd, the generator runs on the path where the integer addition overflowed. B3 assumes that CheckAdd will side-exit upon overflow, so the generator must do some kind of termination. Usually, this is used to implement OSR exits on integer overflow and the optional arguments to CheckAdd will be the OSR exit state. Must use the CheckValue class.

T CheckSub(T, T, [T2, [T3, ...]])

Checked integer subtraction. Works for T = Int32 or T = Int64. First first two children are mandatory. Additional children are optional. All of the Check instructions take a generator and value constraints like a Patchpoint. In the case of a CheckSub, the generator runs on the path where the integer subtraction overflowed. B3 assumes that CheckSub will side-exit upon overflow, so the generator must do some kind of termination. Usually, this is used to implement OSR exits on integer overflow and the optional arguments to CheckSub will be the OSR exit state. You can use CheckSub for checked negation by using zero for the first child. B3 will select the native negate instruction when you do this. Must use the CheckValue class.

T CheckMul(T, T, [T2, [T3, ...]])

Checked integer multiplication. Works for T = Int32 or T = Int64. First first two children are mandatory. Additional children are optional. All of the Check instructions take a generator and value constraints like a Patchpoint. In the case of a CheckMul, the generator runs on the path where the integer multiplication overflowed. B3 assumes that CheckMul will side-exit upon overflow, so the generator must do some kind of termination. Usually, this is used to implement OSR exits on integer overflow and the optional arguments to CheckMul will be the OSR exit state. Must use the CheckValue class.

Void Check(T, [T2, [T3, ...]])

Exit check. Works for T = Int32 or T = Int64. This branches on the first child. If the first child is zero, this just falls through. If it's non-zero, it goes to the exit path generated by the passed generator. Only the first child is mandatory. B3 assumes that Check will side-exit when the first child is non-zero, so the generator must do some kind of termination. Usually, this is used to implement OSR exit checks and the optional arguments to Check will be the OSR exit state. Check supports efficient compare/branch fusion, so you can Check for fairly sophisticated predicates. For example, Check(Equal(LessThan(@a, @b), 0)) where @a and @b are doubles will be generated to an instruction that branches to the exit if @a >= @b or if either @a or @b are NaN. Must use the CheckValue class.

Void WasmBoundsCheck(Int32, pinnedGPR, offset)

Special Wasm opcode. This branches on the first child. If the first child plus the offset produces a Int64 less than to the pinnedGPR this falls through. Otherwise, it branches to the exit path generated by the passed generator. Unlike the Patch/Check family, the generator used by WasmBoundsCheck should be set on the Procedure itself. The GRPReg passed in pinnedGPR must also be marked as pinned by calling the Procedure's pinning API, or it must be InvalidGPR, in which case the out-of-bounds limit is 4GiB. In the later case a maximum parameter can be provided, to further constrain the out-of-bounds limit and help generate smaller immediates. B3 assumes the WasmBoundsCheck will side-exit when it branches, so the generator must do some kind of termination. In Wasm this is used to trap and unwind back to JS. Must use the WasmBoundsCheckValue class.

Void Upsilon(T, ^phi)

B3 uses SSA. SSA requires that each variable gets assigned to only once anywhere in the procedure. But that doesn't work if you have a variable that is supposed to be the result of merging two values along control flow. B3 uses Phi values to represent value merges, just like SSA compilers usually do. But while other SSA compilers represent the inputs to the Phi by listing the control flow edges from which the Phi gets its values, B3 uses the Upsilon value. Each Phi behaves as if it has a memory location associated with it. Executing the Phi behaves like a load from that memory location. Upsilon(@value, ^phi) behaves like a store of @value into the memory location associated with @phi. We say "^phi" when we mean that we are writing to the memory location associated with the Phi. Must use the UpsilonValue class.

T Phi()

Works for any type except Void. Represents a local memory location large enough to hold T. Loads from that memory location. The only way to store to that location is with Upsilon.

Void Jump(takenBlock)

Jumps to takenBlock. This must appear at the end of the basic block. The basic block must have exactly one successor.

Void Branch(T, takenBlock, notTakenBlock)

Works for T = Int32 or T = Int64. Branches on the child. If the child is non-zero, it branches to the takenBlock. Otherwise it branches to the notTakenBlock. Must appear at the end of the basic block. The block must have exactly two successors.

Void Switch(T, cases...)

Works for T = Int32 or T = Int64. Switches on the child. Contains a list of switch cases. Each switch case has an integer constant and a target block. The switch value also contains a fall-through target in case the child has a value that wasn't mentioned in the cases list. Must use the SwitchValue class. Must appear at the end of the basic block. The block must have one successor for each case, plus a successor for the fall-through (default) case. You can manage the successors of a block containing a Switch using API in SwitchValue, like SwitchValue::appendCase() and SwitchValue::setFallThrough().

Void EntrySwitch()

B3 supports multiple procedure entrypoints. The way you create multiple entrypoints is by placing an EntrySwitch at the end of the root block. The root block must then have a successor for each entrypoint. Additionally, you must tell the Procedure how many entrypoints you want. For example:

Procedure proc;
proc.setNumEntrypoints(3);
BasicBlock* root = proc.addBlock();
BasicBlock* firstEntrypoint = proc.addBlock();
BasicBlock* secondEntrypoint = proc.addBlock();
BasicBlock* thirdEntrypoint = proc.addBlock();
root->appendNew<Value>(proc, EntrySwitch, Origin());
root->appendSuccessor(firstEntrypoint);
root->appendSuccessor(secondEntrypoint);
root->appendSuccessor(thirdEntrypoint);

This is the canonical way to use EntrySwitch, however the semantics are flexible enough to allow its use anywhere in the control flow graph. You can have an arbitrary number of EntrySwitches. This flexibility ensures that EntrySwitch works even when B3 does transformations that move code above the EntrySwitch, duplicate the EntrySwitch itself, or do any number of other unexpected things.

EntrySwitch behaves as if each Procedure has a variable called Entry. Each physical entrypoint sets Entry to the index of that entrypoint (so 0, 1, or 2, above) and jumps to the root block. EntrySwitch is just a switch on the Entry variable. Transformations over code that uses EntrySwitch are valid so long as they don't change the procedure's behavior under these semantics.

EntrySwitch is implemented without any actual variables or switching. Instead, all code that lies on some path from the root block to some EntrySwitch is cloned for each entrypoint. This lowering is done as a very late phase in Air, so most of the compiler does not have to know anything about entrypoints. Most of the compiler treats EntrySwitch as an opaque control flow construct. EntrySwitch lowering is allowed to clone an arbitrary amount of code. However, normal uses of EntrySwitch will place it at the end of an empty root block and B3 will only hoist a handful of things above EntrySwitch. This leads to only a small amount of cloned code in practice.

Void Return(T (optional))

Returns the control flow to the caller and terminates the procedure. Must appear at the end of the basic block. The block must have zero successors.

If the node has a child, its value is returned. The type of the child can be any type except Void.

Void Oops()

Indicates unreachable code. This may be implemented as either a trap or as a bare fall-through, since B3 is allowed to assume that this will never be reached. Must appear at the end of the basic block. The block must have zero successors. Note that we also use the Oops opcode to mean "no such opcode" in some B3 transformations.

Flags

This section describes flags. These may be set in Kind.

Chill: Applies to: Div, Mod. You can create a chill Div/Mod by saying chill(Div). This creates a Kind that has the Chill flag set. This can only be used with interer types. An operation is said to be chill if it returns a sensible value whenever its non-chill form would have had undefined behavior. Chill Div turns x/0 into 0 and -2³¹/-1 into -2³¹. We recognize this in IR because it's exactly the semantics of division on ARM64, and it's also exactly the semantics that JavaScript wants for "(x/y)|0". Chill Mod turns x%0 into 0 and -2³¹%-1 into 0. This matches the semantics of JavaScript "(x%y)|0".
Traps: Applies to: Load8Z, Load8S, Load16Z, Load16S, Load, Store8, Store16, Store. You can create a trapping Kind from an opcode by saying trapping(opcode). For example, trapping(Load). An operation is said to be trapping if it will cause a page fault when used on an invalid pointer and this page fault will be observed. This means that the compiler cannot fudge when the page fault happens. This is logically equivalent to what B3 calls Effects::exitsSideways, but further implies that if any of the B3 values used to fuse an Air instruction were trapping, then the Air instruction must have its Air::Kind::traps flag set. The compiler won't help you identify where you trapped. Even if you use the compiler's origin facility to track down the trap location, you may get the origin of any B3 value that was incorporated into the fused instruction that caused the trap. For example, "Add(Load<Traps>(ptr), $1)" may claim to trap at the Add rather than the Load on x86, because this pattern is a perfect candidate for add-load fusion. Nevertheless, you are guaranteed to get the trap and the trap will be observed at the point you intended. For example, the compiler will not hoist a trapping load past any effects, even those outside of its read range, because the trap is presumed to read top. The compiler will not attempt to DCE a trapping load. The compiler will not attempt to sink or eliminate any trapping stores, even if they are dead because of a guaranteed subsequent store to the same address, because we conservatively assume that the store was done for the trap effect. This feature is meant to support high throughput memory safety checks in WebAssembly.