Skip to content

Latest commit

 

History

History
702 lines (499 loc) · 24.7 KB

GlobalISel.rst

File metadata and controls

702 lines (499 loc) · 24.7 KB

Global Instruction Selection

Warning

This document is a work in progress. It reflects the current state of the implementation, as well as open design and implementation issues.

GlobalISel is a framework that provides a set of reusable passes and utilities for instruction selection --- translation from LLVM IR to target-specific Machine IR (MIR).

GlobalISel is intended to be a replacement for SelectionDAG and FastISel, to solve three major problems:

  • Performance --- SelectionDAG introduces a dedicated intermediate representation, which has a compile-time cost.

    GlobalISel directly operates on the post-isel representation used by the rest of the code generator, MIR. It does require extensions to that representation to support arbitrary incoming IR: :ref:`gmir`.

  • Granularity --- SelectionDAG and FastISel operate on individual basic blocks, losing some global optimization opportunities.

    GlobalISel operates on the whole function.

  • Modularity --- SelectionDAG and FastISel are radically different and share very little code.

    GlobalISel is built in a way that enables code reuse. For instance, both the optimized and fast selectors share the :ref:`pipeline`, and targets can configure that pipeline to better suit their needs.

Machine IR operates on physical registers, register classes, and (mostly) target-specific instructions.

To bridge the gap with LLVM IR, GlobalISel introduces "generic" extensions to Machine IR:

NOTE: The generic MIR (GMIR) representation still contains references to IR constructs (such as GlobalValue). Removing those should let us write more accurate tests, or delete IR after building the initial MIR. However, it is not part of the GlobalISel effort.

The main addition is support for pre-isel generic machine instructions (e.g., G_ADD). Like other target-independent instructions (e.g., COPY or PHI), these are available on all targets.

TODO: While we're progressively adding instructions, one kind in particular exposes interesting problems: compares and how to represent condition codes. Some targets (x86, ARM) have generic comparisons setting multiple flags, which are then used by predicated variants. Others (IR) specify the predicate in the comparison and users just get a single bit. SelectionDAG uses SETCC/CONDBR vs BR_CC (and similar for select) to represent this.

The MachineIRBuilder class wraps the MachineInstrBuilder and provides a convenient way to create these generic instructions.

Generic instructions operate on a new kind of register: "generic" virtual registers. As opposed to non-generic vregs, they are not assigned a Register Class. Instead, generic vregs have a :ref:`gmir-llt`, and can be assigned a :ref:`gmir-regbank`.

MachineRegisterInfo tracks the same information that it does for non-generic vregs (e.g., use-def chains). Additionally, it also tracks the :ref:`gmir-llt` of the register, and, instead of the TargetRegisterClass, its :ref:`gmir-regbank`, if any.

For simplicity, most generic instructions only accept generic vregs:

  • instead of immediates, they use a gvreg defined by an instruction materializing the immediate value (see :ref:`irtranslator-constants`).
  • instead of physical register, they use a gvreg defined by a COPY.

NOTE: We started with an alternative representation, where MRI tracks a size for each gvreg, 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. A bank has a size, which is the maximum store size of all covered classes.

In general, cross-class copies inside a bank are expected to be cheaper than copies across banks. They are also coalesceable by the register coalescer, whereas cross-bank copies are not.

Also, equivalent operations can be performed on different banks using different instructions.

For example, 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).

Register banks are described by a target-provided API, :ref:`RegisterBankInfo <api-registerbankinfo>`.

Additionally, every generic virtual register has a type, represented by an instance of the LLT class.

Like EVT/MVT/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:

  • sN for scalars
  • pN for pointers
  • <N x sM> for vectors
  • unsized for labels, etc..

LLT is intended to replace the usage of EVT in SelectionDAG.

Here are some LLT examples and their EVT and Type equivalents:

LLT EVT IR Type
s1 i1 i1
s8 i8 i8
s32 i32 i32
s32 f32 float
s17 i17 i17
s16 N/A {i8, i8}
s32 N/A [4 x i8]
p0 iPTR i8*, i32*, %opaque*
p2 iPTR i8 addrspace(2)*
<4 x s32> v4f32 <4 x float>
s64 v1f64 <1 x double>
<3 x s32> v3i32 <3 x i32>
unsized Other label

Rationale: instructions already encode a specific interpretation of types (e.g., add vs. fadd, or sdiv vs. 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.

NOTE: 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.

TODO: Currently, non-generic virtual registers, defined by non-pre-isel-generic instructions, cannot have a type, and thus cannot be used by a pre-isel generic instruction. Instead, they are given a type using a COPY. We could relax that and allow types on all vregs: this would reduce the number of MI required when emitting target-specific MIR early in the pipeline. This should purely be a compile-time optimization.

There are four required passes, regardless of the optimization mode:

Additional passes can then be inserted at higher optimization levels or for specific targets. For example, to match the current SelectionDAG set of transformations: MachineCSE and a better MachineCombiner between every pass.

NOTE: In theory, not all passes are always necessary. As an additional compile-time optimization, we could skip some of the passes by setting the relevant MachineFunction properties. For instance, if the IRTranslator did not encounter any illegal instruction, it would set the legalized property to avoid running the :ref:`milegalizer`. Similarly, we considered specializing the IRTranslator per-target to directly emit target-specific MI. However, we instead decided to keep the core pipeline simple, and focus on minimizing the overhead of the passes in the no-op cases.

This pass translates the input LLVM IR Function to a GMIR MachineFunction.

TODO: This currently doesn't support the more complex instructions, in particular those involving control flow (switch, invoke, ...). For switch in particular, we can initially use the LowerSwitch pass.

The IRTranslator (using the CallLowering target-provided utility) also implements the ABI's calling convention by lowering calls, returns, and arguments to the appropriate physical register usage and instruction sequences.

Aggregates are lowered to a single scalar vreg. This differs from SelectionDAG's multiple vregs via GetValueVTs.

TODO: As some of the bits are undef (padding), we should consider augmenting the representation with additional metadata (in effect, caching computeKnownBits information on vregs). See PR26161: [GlobalISel] Value to vreg during IR to MachineInstr translation for aggregate type

The IRTranslator lowers Constant operands into uses of gvregs defined by G_CONSTANT or G_FCONSTANT instructions. Currently, these instructions are always emitted in the entry basic block. In a MachineFunction, each Constant is materialized by a single gvreg.

This is beneficial as it allows us to fold constants into immediate operands during :ref:`instructionselect`, while still avoiding redundant materializations for expensive non-foldable constants. However, this can lead to unnecessary spills and reloads in an -O0 pipeline, as these vregs can have long live ranges.

TODO: We're investigating better placement of these instructions, in fast and optimized modes.

This pass transforms the generic machine instructions such that they are legal.

A legal instruction is defined as:

  • selectable --- the target will later be able to select it to a target-specific (non-generic) instruction.
  • operating on vregs that can be loaded and stored -- if necessary, the target can select a G_LOAD/G_STORE of each gvreg operand.

As opposed to SelectionDAG, there are no legalization phases. In particular, 'type' and 'operation' legalization are not separate.

Legalization is iterative, and all state is contained in GMIR. To maintain the validity of the intermediate code, instructions are introduced:

  • G_MERGE_VALUES --- concatenate multiple registers of the same size into a single wider register.
  • G_UNMERGE_VALUES --- extract multiple registers of the same size from a single wider register.
  • G_EXTRACT --- extract a simple register (as contiguous sequences of bits) from a single wider register.

As they are expected to be temporary byproducts of the legalization process, they are combined at the end of the :ref:`milegalizer` pass. If any remain, they are expected to always be selectable, using loads and stores if necessary.

Currently the API is broadly similar to SelectionDAG/TargetLowering, but extended in two ways:

  • The set of available actions is wider, avoiding the currently very overloaded Expand (which can cover everything from libcalls to scalarization depending on the node's opcode).
  • Since there's no separate type legalization, independently varying types on an instruction can have independent actions. For example a G_ICMP has 2 independent types: the result and the inputs; we need to be able to say that comparing 2 s32s is OK, but the s1 result must be dealt with in another way.

As such, the primary key when deciding what to do is the InstrAspect, essentially a tuple consisting of (Opcode, TypeIdx, Type) and mapping to a suggested course of action.

An example use might be:

// The CPU can't deal with an s1 result, do something about it.
setAction({G_ICMP, 0, s1}, WidenScalar);
// An s32 input (the second type) is fine though.
setAction({G_ICMP, 1, s32}, Legal);

TODO: An alternative worth investigating is to generalize the API to represent actions using std::function that implements the action, instead of explicit enum tokens (Legal, WidenScalar, ...).

TODO: Moreover, we could use TableGen to initially infer legality of operation from existing patterns (as any pattern we can select is by definition legal). Expanding that to describe legalization actions is a much larger but potentially useful project.

TODO: Types which have a size that isn't a power of 2 aren't currently supported. The setAction API will probably require changes to support them. Even notionally explicitly specified operations only make suggestions like "Widen" or "Narrow". The eventual type is still unspecified and a search is performed by repeated doubling/halving of the type's size. This is incorrect for types that aren't a power of 2. It's reasonable to expect we could construct an efficient set of side-tables for more general lookups though, encoding a map from the integers (i.e. the size of the current type) to types (the legal size).

Vectors first get their element type legalized: <A x sB> becomes <A x sC> such that at least one operation is legal with sC.

This is currently specified by the function setScalarInVectorAction, called for example as:

setScalarInVectorAction(G_ICMP, s1, WidenScalar);

Next the number of elements is chosen so that the entire operation is legal. This aspect is not controllable at the moment, but probably should be (you could imagine disagreements on whether a <2 x s8> operation should be scalarized or extended to <8 x s8>).

This pass constrains the :ref:`gmir-gvregs` operands of generic instructions to some :ref:`gmir-regbank`.

It iteratively maps instructions to a set of per-operand bank assignment. The possible mappings are determined by the target-provided :ref:`RegisterBankInfo <api-registerbankinfo>`. The mapping is then applied, possibly introducing COPY instructions if necessary.

It traverses the MachineFunction top down so that all operands are already mapped when analyzing an instruction.

This pass could also remap target-specific instructions when beneficial. In the future, this could replace the ExeDepsFix pass, as we can directly select the best variant for an instruction that's available on multiple banks.

The RegisterBankInfo class describes multiple aspects of register banks.

  • Banks: addRegBankCoverage --- which register bank covers each register class.
  • Cross-Bank Copies: copyCost --- the cost of a COPY from one bank to another.
  • Default Mapping: getInstrMapping --- the default bank assignments for a given instruction.
  • Alternative Mapping: getInstrAlternativeMapping --- the other possible bank assignments for a given instruction.

TODO: All this information should eventually be static and generated by TableGen, mostly using existing information augmented by bank descriptions.

TODO: getInstrMapping is currently separate from getInstrAlternativeMapping because the latter is more expensive: as we move to static mapping info, both methods should be free, and we should merge them.

RegBankSelect currently has two modes:

  • Fast --- For each instruction, pick a target-provided "default" bank assignment. This is the default at -O0.
  • Greedy --- For each instruction, pick the cheapest of several target-provided bank assignment alternatives.

We intend to eventually introduce an additional optimizing mode:

  • Global --- Across multiple instructions, pick the cheapest combination of bank assignments.

NOTE: On AArch64, we are considering using the Greedy mode even at -O0 (or perhaps at backend -O1): because :ref:`gmir-llt` doesn't distinguish floating point from integer scalars, the default assignment for loads and stores is the integer bank, introducing cross-bank copies on most floating point operations.

This pass transforms generic machine instructions into equivalent target-specific instructions. It traverses the MachineFunction bottom-up, selecting uses before definitions, enabling trivial dead code elimination.

The target implements the InstructionSelector class, containing the target-specific selection logic proper.

The instance is provided by the subtarget, so that it can specialize the selector by subtarget feature (with, e.g., a vector selector overriding parts of a general-purpose common selector). We might also want to parameterize it by MachineFunction, to enable selector variants based on function attributes like optsize.

The simple API consists of:

virtual bool select(MachineInstr &MI)

This target-provided method is responsible for mutating (or replacing) a possibly-generic MI into a fully target-specific equivalent. It is also responsible for doing the necessary constraining of gvregs into the appropriate register classes as well as passing through COPY instructions to the register allocator.

The InstructionSelector can fold other instructions into the selected MI, by walking the use-def chain of the vreg operands. As GlobalISel is Global, this folding can occur across basic blocks.

TableGen will import SelectionDAG rules and provide the following function to execute them:

bool selectImpl(MachineInstr &MI)

The --stats option can be used to determine what proportion of rules were successfully imported. The easiest way to use this is to copy the -gen-globalisel tablegen command from ninja -v and modify it.

Similarly, the --warn-on-skipped-patterns option can be used to obtain the reasons that rules weren't imported. This can be used to focus on the most important rejection reasons.

PatLeafs cannot be imported because their C++ is implemented in terms of SDNode objects. PatLeafs that handle immediate predicates should be replaced by ImmLeaf, IntImmLeaf, or FPImmLeaf as appropriate.

There's no standard answer for other PatLeafs. Some standard predicates have been baked into TableGen but this should not generally be done.

Custom SDNodes should be mapped to Target Pseudos using GINodeEquiv. This will cause the instruction selector to import them but you will also need to ensure the target pseudo is introduced to the MIR before the instruction selector. Any preceeding pass is suitable but the legalizer will be a particularly common choice.

ComplexPatterns cannot be imported because their C++ is implemented in terms of SDNode objects. GlobalISel versions should be defined with GIComplexOperandMatcher and mapped to ComplexPattern with GIComplexPatternEquiv.

The following predicates are useful for porting ComplexPattern:

  • isBaseWithConstantOffset() - Check for base+offset structures
  • isOperandImmEqual() - Check for a particular constant
  • isObviouslySafeToFold() - Check for reasons an instruction can't be sunk and folded into another.

There are some important points for the C++ implementation:

  • Don't modify MIR in the predicate
  • Renderer lambdas should capture by value to avoid use-after-free. They will be used after the predicate returns.
  • Only create instructions in a renderer lambda. GlobalISel won't clean up things you create but don't use.

Iterative Transformations

Passes are split into small, iterative transformations, with all state represented in the MIR.

This differs from SelectionDAG (in particular, the legalizer) using various in-memory side-tables.

MIR Serialization

:ref:`gmir` is serializable (see :doc:`MIRLangRef`). Combined with :ref:`maintainability-iterative`, this enables much finer-grained testing, rather than requiring large and fragile IR-to-assembly tests.

The current "stage" in the :ref:`pipeline` is represented by a set of MachineFunctionProperties:

  • legalized
  • regBankSelected
  • selected

MachineVerifier

The pass approach lets us use the MachineVerifier to enforce invariants. For instance, a regBankSelected function may not have gvregs without a bank.

TODO: The MachineVerifier being monolithic, some of the checks we want to do can't be integrated to it: GlobalISel is a separate library, so we can't directly reference it from CodeGen. For instance, legality checks are currently done in RegBankSelect/InstructionSelect proper. We could #ifdef out the checks, or we could add some sort of verifier API.

The initial goal is to replace FastISel on AArch64. The next step will be to replace SelectionDAG as the optimized ISel.

NOTE: While we iterate on GlobalISel, we strive to avoid affecting the performance of SelectionDAG, FastISel, or the other MIR passes. For instance, the types of :ref:`gmir-gvregs` are stored in a separate table in MachineRegisterInfo, that is destroyed after :ref:`instructionselect`.

FastISel Replacement

For the initial FastISel replacement, we intend to fallback to SelectionDAG on selection failures.

Currently, compile-time of the fast pipeline is within 1.5x of FastISel. We're optimistic we can get to within 1.1/1.2x, but beating FastISel will be challenging given the multi-pass approach. Still, supporting all IR (via a complete legalizer) and avoiding the fallback to SelectionDAG in the worst case should enable better amortized performance than SelectionDAG+FastISel.

NOTE: We considered never having a fallback to SelectionDAG, instead deciding early whether a given function is supported by GlobalISel or not. The decision would be based on :ref:`milegalizer` queries. We abandoned that for two reasons: a) on IR inputs, we'd need to basically simulate the :ref:`irtranslator`; b) to be robust against unforeseen failures and to enable iterative improvements.

Support For Other Targets

In parallel, we're investigating adding support for other - ideally quite different - targets. For instance, there is some initial AMDGPU support.

There are four major classes to implement by the target:

Additionally:

  • TargetPassConfig --- create the passes constituting the pipeline, including additional passes not included in the :ref:`pipeline`.