Generic Machine IR¶
Generic MIR (gMIR) is an intermediate representation that shares the same data structures as MachineIR (MIR) but has more relaxed constraints. As the compilation pipeline proceeds, these constraints are gradually tightened until gMIR has become MIR.
The rest of this document will assume that you are familiar with the concepts in MachineIR (MIR) and will highlight the differences between MIR and gMIR.
This section expands on Machine Instructions from the MIR Language Reference.
Whereas MIR deals largely in Target Instructions and only has a small set of
target independent opcodes such as
gMIR defines a rich collection of
Generic Opcodes which are target
independent and describe operations which are typically supported by targets.
One example is
G_ADD which is the generic opcode for an integer addition.
More information on each of the generic opcodes can be found at
MachineIRBuilder class wraps the
MachineInstrBuilder and provides
a convenient way to create these generic instructions.
This section expands on Registers from the MIR Language Reference.
Generic virtual registers are like virtual registers but they are not assigned a Register Class constraint. Instead, generic virtual registers have less strict constraints starting with a Low Level Type and then further constrained to a Register Bank. Eventually they will be constrained to a register class at which point they become normal virtual registers.
Generic virtual registers can be used with all the virtual register API’s
MachineRegisterInfo. In particular, the def-use chain API’s can
be used without needing to distinguish them from non-generic virtual registers.
For simplicity, most generic instructions only accept virtual registers (both generic and non-generic). There are some exceptions to this but in general:
instead of immediates, they use a generic virtual register defined by an instruction that materializes the immediate value (see Translation of Constants). Typically this is a G_CONSTANT or a G_FCONSTANT. One example of an exception to this rule is G_SEXT_INREG where having an immediate is mandatory.
instead of physical register, they use a generic virtual register that is either defined by a
COPYfrom the physical register or used by a
COPYthat defines the physical register.
We started with an alternative representation, where MRI tracks a size for each generic virtual register, and instructions have lists of types. That had two flaws: the type and size are redundant, and there was no generic way of getting a given operand’s type (as there was no 1:1 mapping between instruction types and operands). We considered putting the type in some variant of MCInstrDesc instead: See PR26576: [GlobalISel] Generic MachineInstrs need a type but this increases the memory footprint of the related objects
A Register Bank is a set of register classes defined by the target. This definition is rather loose so let’s talk about what they can achieve.
Suppose we have a processor that has two register files, A and B. These are equal in every way and support the same instructions for the same cost. They’re just physically stored apart and each instruction can only access registers from A or B but never a mix of the two. If we want to perform an operation on data that’s in split between the two register files, we must first copy all the data into a single register file.
Given a processor like this, we would benefit from clustering related data together into one register file so that we minimize the cost of copying data back and forth to satisfy the (possibly conflicting) requirements of all the instructions. Register Banks are a means to constrain the register allocator to use a particular register file for a virtual register.
In practice, register files A and B are rarely equal. They can typically store the same data but there’s usually some restrictions on what operations you can do on each register file. A fairly common pattern is for one of them to be accessible to integer operations and the other accessible to floating point operations. To accommodate this, let’s rename A and B to GPR (general purpose registers) and FPR (floating point registers).
We now have some additional constraints that limit us. An operation like G_FMUL has to happen in FPR and G_ADD has to happen in GPR. However, even though this prescribes a lot of the assignments we still have some freedom. A G_LOAD can happen in both GPR and FPR, and which we want depends on who is going to consume the loaded data. Similarly, G_FNEG can happen in both GPR and FPR. If we assign it to FPR, then we’ll use floating point negation. However, if we assign it to GPR then we can equivalently G_XOR the sign bit with 1 to invert it.
In summary, Register Banks are a means of disambiguating between seemingly equivalent choices based on some analysis of the differences when each choice is applied in a given context.
To give some concrete examples:
AArch64 has three main banks. GPR for integer operations, FPR for floating point and also for the NEON vector instruction set. The third is CCR and describes the condition code register used for predication.
MIPS has five main banks of which many programs only really use one or two. GPR is the general purpose bank for integer operations. FGR or CP1 is for the floating point operations as well as the MSA vector instructions and a few other application specific extensions. CP0 is for system registers and few programs will use it. CP2 and CP3 are for any application specific coprocessors that may be present in the chip. Arguably, there is also a sixth for the LO and HI registers but these are only used for the result of a few operations and it’s of questionable value to model distinctly from GPR.
X86 can be seen as having 3 main banks: general-purpose, x87, and vector (which could be further split into a bank per domain for single vs double precision instructions). It also looks like there’s arguably a few more potential banks such as one for the AVX512 Mask Registers.
Register banks are described by a target-provided API, RegisterBankInfo.
Additionally, every generic virtual register has a type, represented by an
instance of the
Type, it has no distinction between unsigned and signed
integer types. Furthermore, it also has no distinction between integer and
floating-point types: it mainly conveys absolutely necessary information, such
as size and number of vector lanes:
<N x sM>for vectors
LLT is intended to replace the usage of
EVT in SelectionDAG.
Here are some LLT examples and their
Rationale: instructions already encode a specific interpretation of types
udiv). Also encoding that
information in the type system requires introducing bitcast with no real
advantage for the selector.
Pointer types are distinguished by address space. This matches IR, as opposed to SelectionDAG where address space is an attribute on operations. This representation better supports pointers having different sizes depending on their addressspace.
Is this still true? I thought we’d removed the 1-element vector concept. Hypothetically, it could be distinct from a scalar but I think we failed to find a real occurrence.
Currently, LLT requires at least 2 elements in vectors, but some targets have the concept of a ‘1-element vector’. Representing them as their underlying scalar type is a nice simplification.
The Generic Opcodes that are available are described at Generic Opcodes.