/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
* vim: set ts=8 sts=2 et sw=2 tw=80:
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
///////////////////////////////////////////////////////////////////////////////
// //
// Documentation. Code starts about 670 lines down from here. //
// //
///////////////////////////////////////////////////////////////////////////////
// [SMDOC] An overview of Ion's register allocator
//
// The intent of this documentation is to give maintainers a map with which to
// navigate the allocator. As further understanding is obtained, it should be
// added to this overview.
//
// Where possible, invariants are stated and are marked "(INVAR)". Many
// details are omitted because their workings are currently unknown. In
// particular, this overview doesn't explain how Intel-style "modify" (tied)
// operands are handled. Facts or invariants that are speculative -- believed
// to be true, but not verified at the time of writing -- are marked "(SPEC)".
//
// The various concepts are interdependent, so a single forwards reading of the
// following won't make much sense. Many concepts are explained only after
// they are mentioned.
//
// Where possible examples are shown. Without those the description is
// excessively abstract.
//
// Names of the form ::name mean BacktrackingAllocator::name.
//
// The description falls into two sections:
//
// * Section 1: A tour of the data structures
// * Section 2: The core allocation loop, and bundle splitting
//
// The allocator sometimes produces poor allocations, with excessive spilling
// and register-to-register moves (bugs 1752520, bug 1714280 bug 1746596).
// Work in bug 1752582 shows we can get better quality allocations from this
// framework without having to make any large (conceptual) changes, by having
// better splitting heuristics.
//
// At https://bugzilla.mozilla.org/show_bug.cgi?id=1758274#c17
// (https://bugzilla.mozilla.org/attachment.cgi?id=9288467) is a document
// written at the same time as these comments. It describes some improvements
// we could make to our splitting heuristics, particularly in the presence of
// loops and calls, and shows why the current implementation sometimes produces
// excessive spilling. It builds on the commentary in this SMDOC.
//
//
// Top level pipeline
// ~~~~~~~~~~~~~~~~~~
// There are three major phases in allocation. They run sequentially, at a
// per-function granularity.
//
// (1) Liveness analysis and bundle formation
// (2) Bundle allocation and last-chance allocation
// (3) Rewriting the function to create MoveGroups and to "install"
// the allocation
//
// The input language (LIR) is in SSA form, and phases (1) and (3) depend on
// that SSAness. Without it the allocator wouldn't work.
//
// The top level function is ::go. The phases are divided into functions as
// follows:
//
// (1) ::buildLivenessInfo, ::mergeAndQueueRegisters
// (2) ::processBundle, ::tryAllocatingRegistersForSpillBundles,
// ::pickStackSlots
// (3) ::createMoveGroupsFromLiveRangeTransitions, ::installAllocationsInLIR,
// ::populateSafepoints, ::annotateMoveGroups
//
// The code in this file is structured as much as possible in the same sequence
// as flow through the pipeline. Hence, top level function ::go is right at
// the end. Where a function depends on helper function(s), the helpers appear
// first.
//
//
// ========================================================================
// ==== ====
// ==== Section 1: A tour of the data structures ====
// ==== ====
// ========================================================================
//
// Here are the key data structures necessary for understanding what follows.
//
// Some basic data structures
// ~~~~~~~~~~~~~~~~~~~~~~~~~~
//
// CodePosition
// ------------
// A CodePosition is an unsigned 32-bit int that indicates an instruction index
// in the incoming LIR. Each LIR actually has two code positions, one to
// denote the "input" point (where, one might imagine, the operands are read,
// at least useAtStart ones) and the "output" point, where operands are
// written. Eg:
//
// Block 0 [successor 2] [successor 1]
// 2-3 WasmParameter [def v1<g>:r14]
// 4-5 WasmCall [use v1:F:r14]
// 6-7 WasmLoadTls [def v2<o>] [use v1:R]
// 8-9 WasmNullConstant [def v3<o>]
// 10-11 Compare [def v4<i>] [use v2:R] [use v3:A]
// 12-13 TestIAndBranch [use v4:R]
//
// So for example the WasmLoadTls insn has its input CodePosition as 6 and
// output point as 7. Input points are even numbered, output points are odd
// numbered. CodePositions 0 and 1 never appear, because LIR instruction IDs
// start at 1. Indeed, CodePosition 0 is assumed to be invalid and hence is
// used as a marker for "unusual conditions" in various places.
//
// Phi nodes exist in the instruction stream too. They always appear at the
// start of blocks (of course) (SPEC), but their start and end points are
// printed for the group as a whole. This is to emphasise that they are really
// parallel assignments and that printing them sequentially would misleadingly
// imply that they are executed sequentially. Example:
//
// Block 6 [successor 7] [successor 8]
// 56-59 Phi [def v19<o>] [use v2:A] [use v5:A] [use v13:A]
// 56-59 Phi [def v20<o>] [use v7:A] [use v14:A] [use v12:A]
// 60-61 WasmLoadSlot [def v21<o>] [use v1:R]
// 62-63 Compare [def v22<i>] [use v20:R] [use v21:A]
// 64-65 TestIAndBranch [use v22:R]
//
// See that both Phis are printed with limits 56-59, even though they are
// stored in the LIR array like regular LIRs and so have code points 56-57 and
// 58-59 in reality.
//
// The process of allocation adds MoveGroup LIRs to the function. Each
// incoming LIR has its own private list of MoveGroups (actually, 3 lists; two
// for moves that conceptually take place before the instruction, and one for
// moves after it). Hence the CodePositions for LIRs (the "62-63", etc, above)
// do not change as a result of allocation.
//
// Virtual registers (vregs) in LIR
// --------------------------------
// The MIR from which the LIR is derived, is standard SSA. That SSAness is
// carried through into the LIR (SPEC). In the examples here, LIR SSA names
// (virtual registers, a.k.a. vregs) are printed as "v<number>". v0 never
// appears and is presumed to be a special value, perhaps "invalid" (SPEC).
//
// The allocator core has a type VirtualRegister, but this is private to the
// allocator and not part of the LIR. It carries far more information than
// merely the name of the vreg. The allocator creates one VirtualRegister
// structure for each vreg in the LIR.
//
// LDefinition and LUse
// --------------------
// These are part of the incoming LIR. Each LIR instruction defines zero or
// more values, and contains one LDefinition for each defined value (SPEC).
// Each instruction has zero or more input operands, each of which has its own
// LUse (SPEC).
//
// Both LDefinition and LUse hold both a virtual register name and, in general,
// a real (physical) register identity. The incoming LIR has the real register
// fields unset, except in places where the incoming LIR has fixed register
// constraints (SPEC). Phase 3 of allocation will visit all of the
// LDefinitions and LUses so as to write into the real register fields the
// decisions made by the allocator. For LUses, this is done by overwriting the
// complete LUse with a different LAllocation, for example LStackSlot. That's
// possible because LUse is a child class of LAllocation.
//
// This action of reading and then later updating LDefinition/LUses is the core
// of the allocator's interface to the outside world.
//
// To make visiting of LDefinitions/LUses possible, the allocator doesn't work
// with LDefinition and LUse directly. Rather it has pointers to them
// (VirtualRegister::def_, UsePosition::use_). Hence Phase 3 can modify the
// LIR in-place.
//
// (INVARs, all SPEC):
//
// - The collective VirtualRegister::def_ values should be unique, and there
// should be a 1:1 mapping between the VirtualRegister::def_ values and the
// LDefinitions in the LIR. (So that the LIR LDefinition has exactly one
// VirtualRegister::def_ to track it). But only for the valid LDefinitions.
// If isBogusTemp() is true, the definition is invalid and doesn't have a
// vreg.
//
// - The same for uses: there must be a 1:1 correspondence between the
// CodePosition::use_ values and the LIR LUses.
//
// - The allocation process must preserve these 1:1 mappings. That implies
// (weaker) that the number of VirtualRegisters and of UsePositions must
// remain constant through allocation. (Eg: losing them would mean that some
// LIR def or use would necessarily not get annotated with its final
// allocation decision. Duplicating them would lead to the possibility of
// conflicting allocation decisions.)
//
// Other comments regarding LIR
// ----------------------------
// The incoming LIR is structured into basic blocks and a CFG, as one would
// expect. These (insns, block boundaries, block edges etc) are available
// through the BacktrackingAllocator object. They are important for Phases 1
// and 3 but not for Phase 2.
//
// Phase 3 "rewrites" the input LIR so as to "install" the final allocation.
// It has to insert MoveGroup instructions, but that isn't done by pushing them
// into the instruction array. Rather, each LIR has 3 auxiliary sets of
// MoveGroups (SPEC): two that "happen" conceptually before the LIR, and one
// that happens after it. The rewriter inserts MoveGroups into one of these 3
// sets, and later code generation phases presumably insert the sets (suitably
// translated) into the final machine code (SPEC).
//
//
// Key data structures: LiveRange, VirtualRegister and LiveBundle
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
//
// These three have central roles in allocation. Of them, LiveRange is the
// most central. VirtualRegister is conceptually important throughout, but
// appears less frequently in the allocator code. LiveBundle is important only
// in Phase 2 (where it is central) and at the end of Phase 1, but plays no
// role in Phase 3.
//
// It's important to understand that LiveRange and VirtualRegister correspond
// to concepts visible in the incoming LIR, which is in SSA form. LiveBundle
// by comparison is related to neither the structure of LIR nor its SSA
// properties. Instead, LiveBundle is an essentially adminstrative structure
// used to accelerate allocation and to implement a crude form of
// move-coalescing.
//
// VirtualRegisters and LiveRanges are (almost) static throughout the process,
// because they reflect aspects of the incoming LIR, which does not change.
// LiveBundles by contrast come and go; they are created, but may be split up
// into new bundles, and old ones abandoned.
//
// Each LiveRange is a member of two different lists.
//
// A VirtualRegister (described in detail below) has a vector of LiveRanges that
// it "owns".
//
// A LiveBundle (also described below) has a linked list of LiveRanges that it
// "owns".
//
// Hence each LiveRange is "owned" by one VirtualRegister and one LiveBundle.
// LiveRanges may have their owning LiveBundle changed as a result of
// splitting. By contrast a LiveRange stays with its "owning" VirtualRegister
// for ever.
//
// A few LiveRanges have no VirtualRegister. This is used to implement
// register spilling for calls. Each physical register that's not preserved
// across a call has a small range that covers the call. It is
// ::buildLivenessInfo that adds these small ranges.
//
// Iterating over every VirtualRegister in the system is a common operation and
// is straightforward because (somewhat redundantly?) the LIRGraph knows the
// number of vregs, and more importantly because BacktrackingAllocator::vregs
// is a vector of all VirtualRegisters. By contrast iterating over every
// LiveBundle in the system is more complex because there is no top-level
// registry of them. It is still possible though. See ::dumpLiveRangesByVReg
// and ::dumpLiveRangesByBundle for example code.
//
// LiveRange
// ---------
// Fundamentally, a LiveRange (often written just "range") is a request for
// storage of a LIR vreg for some contiguous sequence of LIRs. A LiveRange
// generally covers only a fraction of a vreg's overall lifetime, so multiple
// LiveRanges are generally needed to cover the whole lifetime.
//
// A LiveRange contains (amongst other things):
//
// * the vreg for which it is for, as a VirtualRegister*
//
// * the range of CodePositions for which it is for, as a LiveRange::Range
//
// * auxiliary information:
//
// - a boolean that indicates whether this LiveRange defines a value for the
// vreg. If so, that definition is regarded as taking place at the first
// CodePoint of the range.
//
// - a linked list of uses of the vreg within this range. Each use is a pair
// of a CodePosition and an LUse*. (INVAR): the uses are maintained in
// increasing order of CodePosition. Multiple uses at the same
// CodePosition are permitted, since that is necessary to represent an LIR
// that uses the same vreg in more than one of its operand slots.
//
// Some important facts about LiveRanges are best illustrated with examples:
//
// v25 75-82 { 75_def:R 78_v25:A 82_v25:A }
//
// This LiveRange is for vreg v25. The value is defined at CodePosition 75,
// with the LIR requiring that it be in a register. It is used twice at
// positions 78 and 82, both with no constraints (A meaning "any"). The range
// runs from position 75 to 82 inclusive. Note however that LiveRange::Range
// uses non-inclusive range ends; hence its .to field would be 83, not 82.
//
// v26 84-85 { 85_v26:R }
//
// This LiveRange is for vreg v26. Here, there's only a single use of it at
// position 85. Presumably it is defined in some other LiveRange.
//
// v19 74-79 { }
//
// This LiveRange is for vreg v19. There is no def and no uses, so at first
// glance this seems redundant. But it isn't: it still expresses a request for
// storage for v19 across 74-79, because Phase 1 regards v19 as being live in
// this range (meaning: having a value that, if changed in this range, would
// cause the program to fail).
//
// Other points:
//
// * (INVAR) Each vreg/VirtualRegister has at least one LiveRange.
//
// * (INVAR) Exactly one LiveRange of a vreg gives a definition for the value.
// All other LiveRanges must consist only of uses (including zero uses, for a
// "flow-though" range as mentioned above). This requirement follows from
// the fact that LIR is in SSA form.
//
// * It follows from this, that the LiveRanges for a VirtualRegister must form
// a tree, where the parent-child relationship is "control flows directly
// from a parent LiveRange (anywhere in the LiveRange) to a child LiveRange
// (start)". The entire tree carries only one value. This is a use of
// SSAness in the allocator which is fundamental: without SSA input, this
// design would not work.
//
// The root node (LiveRange) in the tree must be one that defines the value,
// and all other nodes must only use or be flow-throughs for the value. It's
// OK for LiveRanges in the tree to overlap, providing that there is a unique
// root node -- otherwise it would be unclear which LiveRange provides the
// value.
//
// The function ::createMoveGroupsFromLiveRangeTransitions runs after all
// LiveBundles have been allocated. It visits each VirtualRegister tree in
// turn. For every parent->child edge in a tree, it creates a MoveGroup that
// copies the value from the parent into the child -- this is how the
// allocator decides where to put MoveGroups. There are various other
// details not described here.
//
// * It's important to understand that a LiveRange carries no meaning about
// control flow beyond that implied by the SSA (hence, dominance)
// relationship between a def and its uses. In particular, there's no
// implication that execution "flowing into" the start of the range implies
// that it will "flow out" of the end. Or that any particular use will or
// will not be executed.
//
// * (very SPEC) Indeed, even if a range has a def, there's no implication that
// a use later in the range will have been immediately preceded by execution
// of the def. It could be that the def is executed, flow jumps somewhere
// else, and later jumps back into the middle of the range, where there are
// then some uses.
//
// * Uses of a vreg by a phi node are not mentioned in the use list of a
// LiveRange. The reasons for this are unknown, but it is speculated that
// this is because we don't need to know about phi uses where we use the list
// of positions. See comments on VirtualRegister::usedByPhi_.
//
// * Similarly, a definition of a vreg by a phi node is not regarded as being a
// definition point (why not?), at least as the view of
// LiveRange::hasDefinition_.
//
// * LiveRanges that nevertheless include a phi-defined value have their first
// point set to the first of the block of phis, even if the var isn't defined
// by that specific phi. Eg:
//
// Block 6 [successor 7] [successor 8]
// 56-59 Phi [def v19<o>] [use v2:A] [use v5:A] [use v13:A]
// 56-59 Phi [def v20<o>] [use v7:A] [use v14:A] [use v12:A]
// 60-61 WasmLoadSlot [def v21<o>] [use v1:R]
// 62-63 Compare [def v22<i>] [use v20:R] [use v21:A]
//
// The relevant live range for v20 is
//
// v20 56-65 { 63_v20:R }
//
// Observe that it starts at 56, not 58.
//
// VirtualRegister
// ---------------
// Each VirtualRegister is associated with an SSA value created by the LIR.
// Fundamentally it is a container to hold all of the LiveRanges that together
// indicate where the value must be kept live. The live ranges are stored in a
// vector, VirtualRegister::ranges_. The set of LiveRanges must logically form
// a tree, rooted at the LiveRange which defines the value.
//
// The live ranges are sorted in order of decreasing start point by construction
// after buildLivenessInfo until the main register allocation loops. At that
// point they can become unsorted and this is important for performance because
// it allows the core allocation code to add/remove ranges more efficiently.
// After all bundles are allocated, sortVirtualRegisterRanges is called to
// ensure the ranges are sorted again.
//
// There are various auxiliary fields, most importantly the LIR node and the
// associated LDefinition that define the value.
//
// It is OK, and quite common, for LiveRanges of a VirtualRegister to overlap.
// The effect will be that, in an overlapped area, there are two storage
// locations holding the value. This is OK -- although wasteful of storage
// resources -- because the SSAness means the value must be the same in both
// locations. Hence there's no questions like "which LiveRange holds the most
// up-to-date value?", since it's all just one value anyway.
//
// Note by contrast, it is *not* OK for the LiveRanges of a LiveBundle to
// overlap.
//
// LiveBundle
// ----------
// Similar to VirtualRegister, a LiveBundle is also, fundamentally, a container
// for a set of LiveRanges. The set is stored as a linked list, rooted at
// LiveBundle::ranges_.
//
// However, the similarity ends there:
//
// * The LiveRanges in a LiveBundle absolutely must not overlap. They must
// indicate disjoint sets of CodePositions, and must be stored in the list in
// order of increasing CodePosition. Because of the no-overlap requirement,
// these ranges form a total ordering regardless of whether one uses the
// LiveRange::Range::from_ or ::to_ fields for comparison.
//
// * The LiveRanges in a LiveBundle can otherwise be entirely arbitrary and
// unrelated. They can be from different VirtualRegisters and can have no
// particular mutual significance w.r.t. the SSAness or structure of the
// input LIR.
//
// LiveBundles are the fundamental unit of allocation. The allocator attempts
// to find a single storage location that will work for all LiveRanges in the
// bundle. That's why the ranges must not overlap. If no such location can be
// found, the allocator may decide to split the bundle into multiple smaller
// bundles. Each of those may be allocated independently.
//
// The other really important field is LiveBundle::alloc_, indicating the
// chosen storage location.
//
// Here's an example, for a LiveBundle before it has been allocated:
//
// LB2(parent=none v3 8-21 { 16_v3:A } ## v3 24-25 { 25_v3:F:xmm0 })
//
// LB merely indicates "LiveBundle", and the 2 is the id_ value (see
// below). This bundle has two LiveRanges
//
// v3 8-21 { 16_v3:A }
// v3 24-25 { 25_v3:F:xmm0 }
//
// both of which (coincidentally) are for the same VirtualRegister, v3.The
// second LiveRange has a fixed use in `xmm0`, whilst the first one doesn't
// care (A meaning "any location") so the allocator *could* choose `xmm0` for
// the bundle as a whole.
//
// One might ask: why bother with LiveBundle at all? After all, it would be
// possible to get correct allocations by allocating each LiveRange
// individually, then leaving ::createMoveGroupsFromLiveRangeTransitions to add
// MoveGroups to join up LiveRanges that form each SSA value tree (that is,
// LiveRanges belonging to each VirtualRegister).
//
// There are two reasons:
//
// (1) By putting multiple LiveRanges into each LiveBundle, we can end up with
// many fewer LiveBundles than LiveRanges. Since the cost of allocating a
// LiveBundle is substantially less than the cost of allocating each of its
// LiveRanges individually, the allocator will run faster.
//
// (2) It provides a crude form of move-coalescing. There are situations where
// it would be beneficial, although not mandatory, to have two LiveRanges
// assigned to the same storage unit. Most importantly: (a) LiveRanges
// that form all of the inputs, and the output, of a phi node. (b)
// LiveRanges for the output and first-operand values in the case where we
// are targetting Intel-style instructions.
//
// In such cases, if the bundle can be allocated as-is, then no extra moves
// are necessary. If not (and the bundle is split), then
// ::createMoveGroupsFromLiveRangeTransitions will later fix things up by
// inserting MoveGroups in the right places.
//
// Merging of LiveRanges into LiveBundles is done in Phase 1, by
// ::mergeAndQueueRegisters, after liveness analysis (which generates only
// LiveRanges).
//
// For the bundle mentioned above, viz
//
// LB2(parent=none v3 8-21 { 16_v3:A } ## v3 24-25 { 25_v3:F:xmm0 })
//
// the allocator did not in the end choose to allocate it to `xmm0`. Instead
// it was split into two bundles, LB6 (a "spill parent", or root node in the
// trees described above), and LB9, a leaf node, that points to its parent LB6:
//
// LB6(parent=none v3 8-21 %xmm1.s { 16_v3:A } ## v3 24-25 %xmm1.s { })
// LB9(parent=LB6 v3 24-25 %xmm0.s { 25_v3:F:rax })
//
// Note that both bundles now have an allocation, and that is printed,
// redundantly, for each LiveRange in the bundle -- hence the repeated
// `%xmm1.s` in the lines above. Since all LiveRanges in a LiveBundle must be
// allocated to the same storage location, we never expect to see output like
// this
//
// LB6(parent=none v3 8-21 %xmm1.s { 16_v3:A } ## v3 24-25 %xmm2.s { })
//
// and that is in any case impossible, since a LiveRange doesn't have an
// LAllocation field. Instead it has a pointer to its LiveBundle, and the
// LAllocation lives in the LiveBundle.
//
// For the resulting allocation (LB6 and LB9), all the LiveRanges are use-only
// or flow-through. There are no definitions. But they are all for the same
// VirtualReg, v3, so they all have the same value. An important question is
// where they "get their value from". The answer is that
// ::createMoveGroupsFromLiveRangeTransitions will have to insert suitable
// MoveGroups so that each LiveRange for v3 can "acquire" the value from a
// previously-executed LiveRange, except for the range that defines it. The
// defining LiveRange is not visible here; either it is in some other
// LiveBundle, not shown, or (more likely) the value is defined by a phi-node,
// in which case, as mentioned previously, it is not shown as having an
// explicit definition in any LiveRange.
//
// LiveBundles also have a `SpillSet* spill_` field (see below) and a
// `LiveBundle* spillParent_`. The latter is used to ensure that all bundles
// originating from an "original" bundle share the same spill slot. The
// `spillParent_` pointers can be thought of creating a 1-level tree, where
// each node points at its parent. Since the tree can be only 1-level, the
// following invariant (INVAR) must be maintained:
//
// * if a LiveBundle has a non-null spillParent_ field, then it is a leaf node,
// and no other LiveBundle can point at this one.
//
// * else (it has a null spillParent_ field) it is a root node, and so other
// LiveBundles may point at it.
//
// LiveBundle has a 32-bit `id_` field. This is used for debug printing, and
// makes it easier to see parent-child relationships induced by the
// `spillParent_` pointers. It's also used for sorting live ranges in
// VirtualRegister::sortRanges.
//
// The "life cycle" of LiveBundles is described in Section 2 below.
//
//
// Secondary data structures: SpillSet, Requirement
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
//
// SpillSet
// --------
// A SpillSet is a container for a set of LiveBundles that have been spilled,
// all of which are assigned the same spill slot. The set is represented as a
// vector of points to LiveBundles. SpillSet also contains the identity of the
// spill slot (its LAllocation).
//
// A LiveBundle, if it is to be spilled, has a pointer to the relevant
// SpillSet, and the SpillSet in turn has a pointer back to the LiveBundle in
// its vector thereof. So LiveBundles (that are to be spilled) and SpillSets
// point at each other.
//
// (SPEC) LiveBundles that are not to be spilled (or for which the decision has
// yet to be made, have their SpillSet pointers as null. (/SPEC)
//
// Requirement
// -----------
// Requirements are used transiently during the main allocation loop. It
// summarises the set of constraints on storage location (must be any register,
// must be this specific register, must be stack, etc) for a LiveBundle. This
// is so that the main allocation loop knows what kind of storage location it
// must choose in order to satisfy all of the defs and uses within the bundle.
//
// What Requirement provides is (a) a partially ordered set of locations, and
// (b) a constraint-merging method `merge`.
//
// Requirement needs a rewrite (and, in fact, that has already happened in
// un-landed code in bug 1758274) for the following reasons:
//
// * it's potentially buggy (bug 1761654), although that doesn't currently
// affect us, for reasons which are unclear.
//
// * the partially ordered set has a top point, meaning "no constraint", but it
// doesn't have a corresponding bottom point, meaning "impossible
// constraints". (So it's a partially ordered set, but not a lattice). This
// leads to awkward coding in some places, which would be simplified if there
// were an explicit way to indicate "impossible constraint".
//
//
// Some ASCII art
// ~~~~~~~~~~~~~~
//
// Here's some not-very-good ASCII art that tries to summarise the data
// structures that persist for the entire allocation of a function:
//
// BacktrackingAllocator
// |
// (vregs)
// |
// v
// |
// VirtualRegister -->--(ins)--> LNode
// | | `->--(def)--> LDefinition
// v ^
// | |
// (ranges vector) |
// | |
// | (vreg)
// `--v->--. |
// \ |
// LiveRange -->--b-->-- LiveRange -->--b-->-- NULL
// / |
// ,--b->--' /
// | (bundle)
// ^ /
// | v
// (ranges) /
// | /
// LiveBundle --s-->- LiveBundle
// | \ / |
// | \ / |
// (spill) ^ ^ (spill)
// | \ / |
// v \ / ^
// | (list) |
// \ | /
// `--->---> SpillSet <--'
//
// --b-- LiveRange::next_: links in the list of LiveRanges that belong to
// a LiveBundle
//
// --v-- VirtualRegister::ranges_: vector of LiveRanges that belong to a
// VirtualRegister
//
// --s-- LiveBundle::spillParent: a possible link to my "spill parent bundle"
//
//
// * LiveRange is in the center. Each LiveRange is a member of a linked list,
// the --b-- list, and is also stored in VirtualRegister's `ranges` vector.
//
// * VirtualRegister has a `ranges` vector that contains its LiveRanges.
//
// * LiveBundle has a pointer `ranges` that points to the start of its --b--
// list of LiveRanges.
//
// * LiveRange points back at both its owning VirtualRegister (`vreg`) and its
// owning LiveBundle (`bundle`).
//
// * LiveBundle has a pointer --s-- `spillParent`, which may be null, to its
// conceptual "spill parent bundle", as discussed in detail above.
//
// * LiveBundle has a pointer `spill` to its SpillSet.
//
// * SpillSet has a vector `list` of pointers back to the LiveBundles that
// point at it.
//
// * VirtualRegister has pointers `ins` to the LNode that defines the value and
// `def` to the LDefinition within that LNode.
//
// * BacktrackingAllocator has a vector `vregs` of pointers to all the
// VirtualRegisters for the function. There is no equivalent top-level table
// of all the LiveBundles for the function.
//
// Note that none of these pointers are "owning" in the C++-storage-management
// sense. Rather, everything is stored in single arena which is freed when
// compilation of the function is complete. For this reason,
// BacktrackingAllocator.{h,cpp} is almost completely free of the usual C++
// storage-management artefacts one would normally expect to see.
//
//
// ========================================================================
// ==== ====
// ==== Section 2: The core allocation loop, and bundle splitting ====
// ==== ====
// ========================================================================
//
// Phase 1 of the allocator (described at the start of this SMDOC) computes
// live ranges, merges them into bundles, and places the bundles in a priority
// queue ::allocationQueue, ordered by what ::computePriority computes.
//
//
// The allocation loops
// ~~~~~~~~~~~~~~~~~~~~
// The core of the allocation machinery consists of two loops followed by a
// call to ::pickStackSlots. The latter is uninteresting. The two loops live
// in ::go and are documented in detail there.
//
//
// Bundle splitting
// ~~~~~~~~~~~~~~~~
// If the first of the two abovementioned loops cannot find a register for a
// bundle, either directly or as a result of evicting conflicting bundles, then
// it will have to either split or spill the bundle. The entry point to the
// split/spill subsystem is ::chooseBundleSplit. See comments there.
///////////////////////////////////////////////////////////////////////////////
// //
// End of documentation //
// //
///////////////////////////////////////////////////////////////////////////////
#include "jit/BacktrackingAllocator.h"
#include "mozilla/BinarySearch.h"
#include "mozilla/Maybe.h"
#include <algorithm>
#include "jit/BitSet.h"
#include "jit/CompileInfo.h"
#include "js/Printf.h"
using namespace js;
using namespace js::jit;
using mozilla::DebugOnly;
using mozilla::Maybe;
// This is a big, complex file. Code is grouped into various sections, each
// preceded by a box comment. Sections not marked as "Misc helpers" are
// pretty much the top level flow, and are presented roughly in the same order
// in which the allocation pipeline operates. BacktrackingAllocator::go,
// right at the end of the file, is a good starting point.
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: linked-list management //
// //
///////////////////////////////////////////////////////////////////////////////
static inline bool SortBefore(UsePosition* a, UsePosition* b) {
return a->pos <= b->pos;
}
static inline bool SortBefore(LiveRange* a, LiveRange* b) {
MOZ_ASSERT(!a->intersects(b));
return a->from() < b->from();
}
template <
typename T>
static void InsertSortedList(InlineForwardList<T>& list, T* value,
T* startAt = nullptr) {
if (list.empty()) {
MOZ_ASSERT(!startAt);
list.pushFront(value);
return;
}
#ifdef DEBUG
if (startAt) {
// `startAt` must be an element in `list` that sorts before `value`.
MOZ_ASSERT(SortBefore(startAt, value));
MOZ_ASSERT_IF(*list.begin() == list.back(), list.back() == startAt);
MOZ_ASSERT_IF(startAt != *list.begin(), SortBefore(*list.begin(), startAt));
MOZ_ASSERT_IF(startAt != list.back(), SortBefore(startAt, list.back()));
}
#endif
if (SortBefore(list.back(), value)) {
list.pushBack(value);
return;
}
// If `startAt` is non-nullptr, we can start iterating there.
T* prev = nullptr;
InlineForwardListIterator<T> iter =
startAt ? list.begin(startAt) : list.begin();
if (startAt) {
// `value` must sort after `startAt` so skip to the next element.
MOZ_ASSERT(!SortBefore(value, *iter));
++iter;
prev = startAt;
}
for (; iter; iter++) {
if (SortBefore(value, *iter)) {
break;
}
prev = *iter;
}
if (prev) {
list.insertAfter(prev, value);
}
else {
list.pushFront(value);
}
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: methods for class SpillSet //
// //
///////////////////////////////////////////////////////////////////////////////
void SpillSet::setAllocation(LAllocation alloc) {
for (size_t i = 0; i < numSpilledBundles(); i++) {
spilledBundle(i)->setAllocation(alloc);
}
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: methods for class LiveRange //
// //
///////////////////////////////////////////////////////////////////////////////
static size_t SpillWeightFromUsePolicy(LUse::Policy policy) {
switch (policy) {
case LUse::ANY:
return 1000;
case LUse::
REGISTER:
case LUse::FIXED:
return 2000;
default:
return 0;
}
}
inline void LiveRange::noteAddedUse(UsePosition* use) {
LUse::Policy policy = use->usePolicy();
usesSpillWeight_ += SpillWeightFromUsePolicy(policy);
if (policy == LUse::FIXED) {
++numFixedUses_;
}
}
inline void LiveRange::noteRemovedUse(UsePosition* use) {
LUse::Policy policy = use->usePolicy();
usesSpillWeight_ -= SpillWeightFromUsePolicy(policy);
if (policy == LUse::FIXED) {
--numFixedUses_;
}
MOZ_ASSERT_IF(!hasUses(), !usesSpillWeight_ && !numFixedUses_);
}
void LiveRange::addUse(UsePosition* use) {
MOZ_ASSERT(covers(use->pos));
InsertSortedList(uses_, use);
noteAddedUse(use);
}
UsePosition* LiveRange::popUse() {
UsePosition* ret = uses_.popFront();
noteRemovedUse(ret);
return ret;
}
void LiveRange::tryToMoveDefAndUsesInto(LiveRange* other) {
MOZ_ASSERT(&other->vreg() == &vreg());
MOZ_ASSERT(
this != other);
// This method shouldn't be called for two non-intersecting live ranges
// because it's a no-op in that case.
MOZ_ASSERT(intersects(other));
CodePosition otherFrom = other->from();
CodePosition otherTo = other->to();
// The uses are sorted by position, so first skip all uses before |other|
// starts.
UsePositionIterator iter = usesBegin();
while (iter && iter->pos < otherFrom) {
iter++;
}
// Move over all uses which fit in |other|'s boundaries.
while (iter && iter->pos < otherTo) {
UsePosition* use = *iter;
MOZ_ASSERT(other->covers(use->pos));
uses_.removeAndIncrement(iter);
noteRemovedUse(use);
other->addUse(use);
}
MOZ_ASSERT_IF(iter, !other->covers(iter->pos));
// Distribute the definition to |other| as well, if possible.
if (hasDefinition() && from() == other->from()) {
other->setHasDefinition();
}
}
void LiveRange::moveAllUsesToTheEndOf(LiveRange* other) {
MOZ_ASSERT(&other->vreg() == &vreg());
MOZ_ASSERT(
this != other);
MOZ_ASSERT(other->contains(
this));
if (uses_.empty()) {
return;
}
// Assert |other->uses_| remains sorted after adding our uses at the end.
MOZ_ASSERT_IF(!other->uses_.empty(),
SortBefore(other->uses_.back(), *uses_.begin()));
other->uses_.extendBack(std::move(uses_));
MOZ_ASSERT(!hasUses());
other->usesSpillWeight_ += usesSpillWeight_;
other->numFixedUses_ += numFixedUses_;
usesSpillWeight_ = 0;
numFixedUses_ = 0;
}
bool LiveRange::contains(LiveRange* other)
const {
return from() <= other->from() && to() >= other->to();
}
void LiveRange::intersect(LiveRange* other, Range* pre, Range* inside,
Range* post)
const {
MOZ_ASSERT(pre->empty() && inside->empty() && post->empty());
CodePosition innerFrom = from();
if (from() < other->from()) {
if (to() < other->from()) {
*pre = range_;
return;
}
*pre = Range(from(), other->from());
innerFrom = other->from();
}
CodePosition innerTo = to();
if (to() > other->to()) {
if (from() >= other->to()) {
*post = range_;
return;
}
*post = Range(other->to(), to());
innerTo = other->to();
}
if (innerFrom != innerTo) {
*inside = Range(innerFrom, innerTo);
}
}
bool LiveRange::intersects(LiveRange* other)
const {
Range pre, inside, post;
intersect(other, &pre, &inside, &post);
return !inside.empty();
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: methods for class LiveBundle //
// //
///////////////////////////////////////////////////////////////////////////////
#ifdef DEBUG
size_t LiveBundle::numRanges()
const {
size_t count = 0;
for (RangeIterator iter = rangesBegin(); iter; iter++) {
count++;
}
return count;
}
#endif
LiveRange* LiveBundle::rangeFor(CodePosition pos)
const {
for (RangeIterator iter = rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (range->covers(pos)) {
return range;
}
}
return nullptr;
}
void LiveBundle::addRange(LiveRange* range,
LiveRange* startAt
/* = nullptr */) {
MOZ_ASSERT(!range->bundle());
MOZ_ASSERT(range->hasVreg());
MOZ_ASSERT_IF(startAt, startAt->bundle() ==
this);
range->setBundle(
this);
InsertSortedList(ranges_, range, startAt);
}
void LiveBundle::addRangeAtEnd(LiveRange* range) {
// Note: this method is functionally equivalent to `addRange`, but can be used
// when we know `range` must be added after all existing ranges in the bundle.
MOZ_ASSERT(!range->bundle());
MOZ_ASSERT(range->hasVreg());
MOZ_ASSERT_IF(!ranges_.empty(), SortBefore(ranges_.back(), range));
range->setBundle(
this);
ranges_.pushBack(range);
}
bool LiveBundle::addRangeAtEnd(TempAllocator& alloc, VirtualRegister* vreg,
CodePosition from, CodePosition to) {
LiveRange* range = LiveRange::FallibleNew(alloc, vreg, from, to);
if (!range) {
return false;
}
addRangeAtEnd(range);
return true;
}
bool LiveBundle::addRangeAndDistributeUses(TempAllocator& alloc,
LiveRange* oldRange,
CodePosition from, CodePosition to) {
LiveRange* range = LiveRange::FallibleNew(alloc, &oldRange->vreg(), from, to);
if (!range) {
return false;
}
addRange(range);
oldRange->tryToMoveDefAndUsesInto(range);
return true;
}
LiveRange* LiveBundle::popFirstRange() {
RangeIterator iter = rangesBegin();
if (!iter) {
return nullptr;
}
LiveRange* range = *iter;
ranges_.removeAt(iter);
range->setBundle(nullptr);
return range;
}
void LiveBundle::removeRange(LiveRange* range) {
for (RangeIterator iter = rangesBegin(); iter; iter++) {
LiveRange* existing = *iter;
if (existing == range) {
ranges_.removeAt(iter);
return;
}
}
MOZ_CRASH();
}
void LiveBundle::removeAllRangesFromVirtualRegisters() {
VirtualRegister* prevVreg = nullptr;
for (RangeIterator iter = rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
MOZ_ASSERT(!range->hasUses());
if (&range->vreg() != prevVreg) {
// As an optimization, remove all ranges owned by this bundle from the
// register's list, instead of a single range at a time.
range->vreg().removeRangesForBundle(
this);
prevVreg = &range->vreg();
}
}
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: methods for class VirtualRegister //
// //
///////////////////////////////////////////////////////////////////////////////
bool VirtualRegister::addInitialRange(TempAllocator& alloc, CodePosition from,
CodePosition to) {
MOZ_ASSERT(rangesSorted_,
"ranges stay sorted during live range building");
MOZ_ASSERT(from < to);
MOZ_ASSERT_IF(hasRanges(), from < ranges_.back()->to());
// Mark [from,to) as a live range for this register during the initial
// liveness analysis, coalescing with any existing overlapping ranges.
LiveRange* merged = nullptr;
for (RangeIterator iter(*
this); iter; iter++) {
LiveRange* existing = *iter;
MOZ_ASSERT(from < existing->to());
if (to < existing->from()) {
// All other ranges start after |to| so can't overlap.
break;
}
if (!merged) {
// This is the first old range we've found that overlaps the new
// range. Extend this one to cover its union with the new range.
merged = existing;
if (from < existing->from()) {
existing->setFrom(from);
}
if (to > existing->to()) {
existing->setTo(to);
}
// Continue searching to see if any other old ranges can be
// coalesced with the new merged range.
}
else {
// Coalesce this range into the previous range we merged into.
MOZ_ASSERT(existing->from() >= merged->from());
if (existing->to() > merged->to()) {
merged->setTo(existing->to());
}
MOZ_ASSERT(!existing->hasDefinition());
existing->moveAllUsesToTheEndOf(merged);
MOZ_ASSERT(!existing->hasUses());
}
removeFirstRange(iter);
}
if (merged) {
// Re-append the merged range.
if (!ranges_.append(merged)) {
return false;
}
}
else {
// The new range does not overlap any existing range for the vreg.
MOZ_ASSERT_IF(hasRanges(), to < ranges_.back()->from());
LiveRange* range = LiveRange::FallibleNew(alloc,
this, from, to);
if (!range) {
return false;
}
if (!ranges_.append(range)) {
return false;
}
}
MOZ_ASSERT(rangesSorted_,
"ranges are still sorted");
#ifdef DEBUG
// Check the last few ranges don't overlap and are in the correct order.
size_t len = ranges_.length();
static constexpr size_t MaxIterations = 4;
size_t start = len > MaxIterations ? (len - MaxIterations) : 1;
for (size_t i = start; i < len; i++) {
LiveRange* range = ranges_[i];
LiveRange* prev = ranges_[i - 1];
MOZ_ASSERT(range->from() < range->to());
MOZ_ASSERT(range->to() < prev->from());
}
#endif
return true;
}
void VirtualRegister::addInitialUse(UsePosition* use) {
MOZ_ASSERT(rangesSorted_,
"ranges stay sorted during live range building");
ranges_.back()->addUse(use);
}
void VirtualRegister::setInitialDefinition(CodePosition from) {
MOZ_ASSERT(rangesSorted_,
"ranges stay sorted during live range building");
LiveRange* first = ranges_.back();
MOZ_ASSERT(from >= first->from());
first->setFrom(from);
first->setHasDefinition();
}
LiveRange* VirtualRegister::rangeFor(CodePosition pos,
bool preferRegister
/* = false */) const {
assertRangesSorted();
size_t len = ranges_.length();
// Because the ranges are sorted in order of descending start position, we use
// binary search to find the first range where |range->from <= pos|. All
// ranges before that definitely don't cover |pos|.
auto compare = [pos](LiveRange* other) {
if (pos < other->from()) {
return 1;
}
if (pos > other->from()) {
return -1;
}
return 0;
};
size_t index;
mozilla::BinarySearchIf(ranges_, 0, len, compare, &index);
if (index == len) {
// None of the ranges overlap.
MOZ_ASSERT(ranges_.back()->from() > pos);
return nullptr;
}
// There can be multiple ranges with |range->from == pos|. We want to start at
// the first one.
while (index > 0 && ranges_[index - 1]->from() == pos) {
index--;
}
// Verify the above code is correct:
// * The range at |index| starts before or at |pos| and needs to be searched.
// * All ranges before |index| start after |pos| and can be ignored.
MOZ_ASSERT(ranges_[index]->from() <= pos);
MOZ_ASSERT_IF(index > 0, ranges_[index - 1]->from() > pos);
LiveRange* found = nullptr;
do {
LiveRange* range = ranges_[index];
if (range->covers(pos)) {
if (!preferRegister || range->bundle()->allocation().isRegister()) {
return range;
}
if (!found) {
found = range;
}
}
index++;
}
while (index < len);
return found;
}
void VirtualRegister::sortRanges() {
if (rangesSorted_) {
assertRangesSorted();
return;
}
// Sort ranges by start position in descending order.
//
// It would be correct to only compare the start position, but std::sort is
// not a stable sort and we don't want the order of ranges with the same start
// position to depend on the sort implementation. Use the end position and the
// bundle's id to ensure ranges are always sorted the same way.
auto compareRanges = [](LiveRange* a, LiveRange* b) ->
bool {
if (a->from() != b->from()) {
return a->from() > b->from();
}
if (a->to() != b->to()) {
return a->to() > b->to();
}
// Overlapping live ranges must belong to different bundles.
MOZ_ASSERT_IF(a != b, a->bundle()->id() != b->bundle()->id());
return a->bundle()->id() > b->bundle()->id();
};
std::sort(ranges_.begin(), ranges_.end(), compareRanges);
rangesSorted_ =
true;
}
#ifdef DEBUG
void VirtualRegister::assertRangesSorted()
const {
MOZ_ASSERT(rangesSorted_);
// Assert the last N ranges in the vector are sorted correctly. We don't check
// the whole vector to not slow down debug builds too much.
size_t len = ranges_.length();
static constexpr size_t MaxIterations = 4;
size_t start = len > MaxIterations ? (len - MaxIterations) : 1;
for (size_t i = start; i < len; i++) {
LiveRange* prev = ranges_[i - 1];
LiveRange* range = ranges_[i];
MOZ_ASSERT(range->from() <= prev->from());
// If two ranges have the same |from| position, neither must be defining.
// This ensures the defining range, if any, always comes first after
// sorting.
MOZ_ASSERT_IF(range->from() == prev->from(),
!range->hasDefinition() && !prev->hasDefinition());
}
}
#endif
bool VirtualRegister::addRange(LiveRange* range) {
bool sorted = ranges_.empty() ||
(rangesSorted_ && ranges_.back()->from() >= range->from());
if (!ranges_.append(range)) {
return false;
}
rangesSorted_ = sorted;
return true;
}
void VirtualRegister::removeFirstRange(RangeIterator& iter) {
MOZ_ASSERT(iter.index() == ranges_.length() - 1);
ranges_.popBack();
}
void VirtualRegister::removeRangesForBundle(LiveBundle* bundle) {
auto bundleMatches = [bundle](LiveRange* range) {
return range->bundle() == bundle;
};
ranges_.eraseIf(bundleMatches);
}
template <
typename Pred>
void VirtualRegister::removeRangesIf(Pred&& pred) {
assertRangesSorted();
ranges_.eraseIf([&](LiveRange* range) {
return pred(ranges_, range); });
}
// Helper for ::tryMergeReusedRegister.
bool VirtualRegister::replaceLastRangeLinear(LiveRange* old, LiveRange* newPre,
LiveRange* newPost) {
assertRangesSorted();
// |old| is currently the first range in the Vector (the range with the
// highest start position). This range is being replaced with two new ranges,
// |newPre| and |newPost|. The relative order of these ranges by start
// position is:
//
// old <= newPre <= newPost
//
// To keep the vector sorted by descending start position, |newPost| must
// become the new last range and |newPre| must be inserted after it.
MOZ_ASSERT(ranges_[0] == old);
MOZ_ASSERT(old->from() <= newPre->from());
MOZ_ASSERT(newPre->from() <= newPost->from());
ranges_[0] = newPost;
if (!ranges_.insert(ranges_.begin() + 1, newPre)) {
return false;
}
assertRangesSorted();
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: queries about uses //
// //
///////////////////////////////////////////////////////////////////////////////
static inline LDefinition* FindReusingDefOrTemp(LNode* node,
LAllocation* alloc) {
if (node->isPhi()) {
MOZ_ASSERT(node->toPhi()->numDefs() == 1);
MOZ_ASSERT(node->toPhi()->getDef(0)->policy() !=
LDefinition::MUST_REUSE_INPUT);
return nullptr;
}
LInstruction* ins = node->toInstruction();
for (size_t i = 0; i < ins->numDefs(); i++) {
LDefinition* def = ins->getDef(i);
if (def->policy() == LDefinition::MUST_REUSE_INPUT &&
ins->getOperand(def->getReusedInput()) == alloc) {
return def;
}
}
for (size_t i = 0; i < ins->numTemps(); i++) {
LDefinition* def = ins->getTemp(i);
if (def->policy() == LDefinition::MUST_REUSE_INPUT &&
ins->getOperand(def->getReusedInput()) == alloc) {
return def;
}
}
return nullptr;
}
bool BacktrackingAllocator::isReusedInput(LUse* use, LNode* ins,
bool considerCopy) {
if (LDefinition* def = FindReusingDefOrTemp(ins, use)) {
return considerCopy || !vregs[def->virtualRegister()].mustCopyInput();
}
return false;
}
bool BacktrackingAllocator::isRegisterUse(UsePosition* use, LNode* ins,
bool considerCopy) {
switch (use->usePolicy()) {
case LUse::ANY:
return isReusedInput(use->use(), ins, considerCopy);
case LUse::
REGISTER:
case LUse::FIXED:
return true;
default:
return false;
}
}
bool BacktrackingAllocator::isRegisterDefinition(LiveRange* range) {
if (!range->hasDefinition()) {
return false;
}
VirtualRegister& reg = range->vreg();
if (reg.ins()->isPhi()) {
return false;
}
if (reg.def()->policy() == LDefinition::FIXED &&
!reg.def()->output()->isRegister()) {
return false;
}
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: atomic LIR groups //
// //
///////////////////////////////////////////////////////////////////////////////
// The following groupings contain implicit (invisible to ::buildLivenessInfo)
// value flows, and therefore no split points may be requested inside them.
// This is an otherwise unstated part of the contract between LIR generation
// and the allocator.
//
// (1) (any insn) ; OsiPoint
//
// [Further group definitions and supporting code to come, pending rework
// of the wasm atomic-group situation.]
CodePosition RegisterAllocator::minimalDefEnd(LNode* ins)
const {
// Compute the shortest interval that captures vregs defined by ins.
// Watch for instructions that are followed by an OSI point.
// If moves are introduced between the instruction and the OSI point then
// safepoint information for the instruction may be incorrect.
while (
true) {
LNode* next = insData[ins->id() + 1];
if (!next->isOsiPoint()) {
break;
}
ins = next;
}
return outputOf(ins);
}
///////////////////////////////////////////////////////////////////////////////
// //
// Misc helpers: computation of bundle priorities and spill weights //
// //
///////////////////////////////////////////////////////////////////////////////
size_t BacktrackingAllocator::computePriority(LiveBundle* bundle) {
// The priority of a bundle is its total length, so that longer lived
// bundles will be processed before shorter ones (even if the longer ones
// have a low spill weight). See processBundle().
size_t lifetimeTotal = 0;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
lifetimeTotal += range->to() - range->from();
}
return lifetimeTotal;
}
bool BacktrackingAllocator::minimalDef(LiveRange* range, LNode* ins) {
// Whether this is a minimal range capturing a definition at ins.
return (range->to() <= minimalDefEnd(ins).next()) &&
((!ins->isPhi() && range->from() == inputOf(ins)) ||
range->from() == outputOf(ins));
}
bool BacktrackingAllocator::minimalUse(LiveRange* range, UsePosition* use) {
// Whether this is a minimal range capturing |use|.
LNode* ins = insData[use->pos];
return (range->from() == inputOf(ins)) &&
(range->to() ==
(use->use()->usedAtStart() ? outputOf(ins) : outputOf(ins).next()));
}
bool BacktrackingAllocator::minimalBundle(LiveBundle* bundle,
bool* pfixed) {
// Return true iff |bundle| is a minimal bundle. A minimal bundle is a bundle
// that can't be split further. Minimal bundles have a single (very short)
// range.
//
// If the bundle is minimal we also set the |*pfixed| outparam to indicate
// whether the bundle must be allocated to a fixed register. The value of
// |*pfixed| is undefined if this function returns false.
LiveBundle::RangeIterator iter = bundle->rangesBegin();
LiveRange* range = *iter;
MOZ_ASSERT(range->hasVreg(),
"Call ranges are not added to LiveBundles");
// If a bundle contains multiple ranges, splitAtAllRegisterUses will split
// each range into a separate bundle.
if (++iter) {
return false;
}
if (range->hasDefinition()) {
VirtualRegister& reg = range->vreg();
if (!minimalDef(range, reg.ins())) {
return false;
}
if (pfixed) {
*pfixed = reg.def()->policy() == LDefinition::FIXED &&
reg.def()->output()->isRegister();
}
return true;
}
// Performance optimization: |minimalUse| will only return true for length-1
// or length-2 ranges so if this is a longer range it can't be minimal.
if (range->to() - range->from() > 2) {
#ifdef DEBUG
for (UsePositionIterator iter = range->usesBegin(); iter; iter++) {
MOZ_ASSERT(!minimalUse(range, *iter));
}
#endif
return false;
}
bool fixed =
false, minimal =
false, multiple =
false;
for (UsePositionIterator iter = range->usesBegin(); iter; iter++) {
if (iter != range->usesBegin()) {
multiple =
true;
}
switch (iter->usePolicy()) {
case LUse::FIXED:
if (fixed) {
return false;
}
fixed =
true;
if (minimalUse(range, *iter)) {
minimal =
true;
}
break;
case LUse::
REGISTER:
if (minimalUse(range, *iter)) {
minimal =
true;
}
break;
default:
break;
}
}
// If a range contains a fixed use and at least one other use,
// splitAtAllRegisterUses will split each use into a different bundle.
if (multiple && fixed) {
return false;
}
if (!minimal) {
return false;
}
if (pfixed) {
*pfixed = fixed;
}
return true;
}
size_t BacktrackingAllocator::computeSpillWeight(LiveBundle* bundle) {
// Minimal bundles have an extremely high spill weight, to ensure they
// can evict any other bundles and be allocated to a register.
bool fixed;
if (minimalBundle(bundle, &fixed)) {
return fixed ? 2000000 : 1000000;
}
size_t usesTotal = 0;
fixed =
false;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (range->hasDefinition()) {
VirtualRegister& reg = range->vreg();
if (reg.def()->policy() == LDefinition::FIXED &&
reg.def()->output()->isRegister()) {
usesTotal += 2000;
fixed =
true;
}
else if (!reg.ins()->isPhi()) {
usesTotal += 2000;
}
}
usesTotal += range->usesSpillWeight();
if (range->numFixedUses() > 0) {
fixed =
true;
}
}
// Bundles with fixed uses are given a higher spill weight, since they must
// be allocated to a specific register.
if (testbed && fixed) {
usesTotal *= 2;
}
// Compute spill weight as a use density, lowering the weight for long
// lived bundles with relatively few uses.
size_t lifetimeTotal = computePriority(bundle);
return lifetimeTotal ? usesTotal / lifetimeTotal : 0;
}
size_t BacktrackingAllocator::maximumSpillWeight(
const LiveBundleVector& bundles) {
size_t maxWeight = 0;
for (size_t i = 0; i < bundles.length(); i++) {
maxWeight = std::max(maxWeight, computeSpillWeight(bundles[i]));
}
return maxWeight;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Initialization of the allocator //
// //
///////////////////////////////////////////////////////////////////////////////
// This function pre-allocates and initializes as much global state as possible
// to avoid littering the algorithms with memory management cruft.
bool BacktrackingAllocator::init() {
if (!RegisterAllocator::init()) {
return false;
}
uint32_t numBlocks = graph.numBlockIds();
MOZ_ASSERT(liveIn.empty());
if (!liveIn.growBy(numBlocks)) {
return false;
}
size_t numVregs = graph.numVirtualRegisters();
if (!vregs.initCapacity(numVregs)) {
return false;
}
for (uint32_t i = 0; i < numVregs; i++) {
vregs.infallibleEmplaceBack();
}
// Build virtual register objects.
for (size_t i = 0; i < graph.numBlocks(); i++) {
if (mir->shouldCancel(
"Create data structures (main loop)")) {
return false;
}
LBlock* block = graph.getBlock(i);
for (LInstructionIterator ins = block->begin(); ins != block->end();
ins++) {
if (mir->shouldCancel(
"Create data structures (inner loop 1)")) {
return false;
}
for (size_t j = 0; j < ins->numDefs(); j++) {
LDefinition* def = ins->getDef(j);
if (def->isBogusTemp()) {
continue;
}
vreg(def).init(*ins, def,
/* isTemp = */ false);
}
for (size_t j = 0; j < ins->numTemps(); j++) {
LDefinition* def = ins->getTemp(j);
if (def->isBogusTemp()) {
continue;
}
vreg(def).init(*ins, def,
/* isTemp = */ true);
}
}
for (size_t j = 0; j < block->numPhis(); j++) {
LPhi* phi = block->getPhi(j);
LDefinition* def = phi->getDef(0);
vreg(def).init(phi, def,
/* isTemp = */ false);
}
}
LiveRegisterSet remainingRegisters(allRegisters_.asLiveSet());
while (!remainingRegisters.emptyGeneral()) {
AnyRegister reg = AnyRegister(remainingRegisters.takeAnyGeneral());
registers[reg.code()].allocatable =
true;
}
while (!remainingRegisters.emptyFloat()) {
AnyRegister reg =
AnyRegister(remainingRegisters.takeAnyFloat<RegTypeName::Any>());
registers[reg.code()].allocatable =
true;
}
LifoAlloc* lifoAlloc = mir->alloc().lifoAlloc();
for (size_t i = 0; i < AnyRegister::Total; i++) {
registers[i].reg = AnyRegister::FromCode(i);
registers[i].allocations.setAllocator(lifoAlloc);
}
hotcode.setAllocator(lifoAlloc);
// Partition the graph into hot and cold sections, for helping to make
// splitting decisions. Since we don't have any profiling data this is a
// crapshoot, so just mark the bodies of inner loops as hot and everything
// else as cold.
LBlock* backedge = nullptr;
for (size_t i = 0; i < graph.numBlocks(); i++) {
LBlock* block = graph.getBlock(i);
// If we see a loop header, mark the backedge so we know when we have
// hit the end of the loop. Don't process the loop immediately, so that
// if there is an inner loop we will ignore the outer backedge.
if (block->mir()->isLoopHeader()) {
backedge = block->mir()->backedge()->lir();
}
if (block == backedge) {
LBlock* header = block->mir()->loopHeaderOfBackedge()->lir();
LiveRange* range = LiveRange::FallibleNew(
alloc(), nullptr, entryOf(header), exitOf(block).next());
if (!range || !hotcode.insert(range)) {
return false;
}
}
}
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Liveness analysis //
// //
///////////////////////////////////////////////////////////////////////////////
// Helper for ::buildLivenessInfo
bool BacktrackingAllocator::addInitialFixedRange(AnyRegister reg,
CodePosition from,
CodePosition to) {
LiveRange* range = LiveRange::FallibleNew(alloc(), nullptr, from, to);
if (!range) {
return false;
}
LiveRangePlus rangePlus(range);
return registers[reg.code()].allocations.insert(rangePlus);
}
// Helper for ::buildLivenessInfo
#ifdef DEBUG
// Returns true iff ins has a def/temp reusing the input allocation.
static bool IsInputReused(LInstruction* ins, LUse* use) {
for (size_t i = 0; i < ins->numDefs(); i++) {
if (ins->getDef(i)->policy() == LDefinition::MUST_REUSE_INPUT &&
ins->getOperand(ins->getDef(i)->getReusedInput())->toUse() == use) {
return true;
}
}
for (size_t i = 0; i < ins->numTemps(); i++) {
if (ins->getTemp(i)->policy() == LDefinition::MUST_REUSE_INPUT &&
ins->getOperand(ins->getTemp(i)->getReusedInput())->toUse() == use) {
return true;
}
}
return false;
}
#endif
/*
* This function builds up liveness ranges for all virtual registers
* defined in the function.
*
* The algorithm is based on the one published in:
*
* Wimmer, Christian, and Michael Franz. "Linear Scan Register Allocation on
* SSA Form." Proceedings of the International Symposium on Code Generation
* and Optimization. Toronto, Ontario, Canada, ACM. 2010. 170-79. PDF.
*
* The algorithm operates on blocks ordered such that dominators of a block
* are before the block itself, and such that all blocks of a loop are
* contiguous. It proceeds backwards over the instructions in this order,
* marking registers live at their uses, ending their live ranges at
* definitions, and recording which registers are live at the top of every
* block. To deal with loop backedges, registers live at the beginning of
* a loop gain a range covering the entire loop.
*/
bool BacktrackingAllocator::buildLivenessInfo() {
JitSpew(JitSpew_RegAlloc,
"Beginning liveness analysis");
// The callPositions vector is initialized from index |length - 1| to 0, to
// ensure the call positions are sorted.
if (!callPositions.growByUninitialized(graph.numCallInstructions())) {
return false;
}
size_t prevCallPositionIndex = callPositions.length();
for (size_t i = graph.numBlocks(); i > 0; i--) {
if (mir->shouldCancel(
"Build Liveness Info (main loop)")) {
return false;
}
LBlock* block = graph.getBlock(i - 1);
MBasicBlock* mblock = block->mir();
VirtualRegBitSet& live = liveIn[mblock->id()];
// Propagate liveIn from our successors to us.
for (size_t i = 0; i < mblock->lastIns()->numSuccessors(); i++) {
MBasicBlock* successor = mblock->lastIns()->getSuccessor(i);
// Skip backedges, as we fix them up at the loop header.
if (mblock->id() < successor->id()) {
if (!live.insertAll(liveIn[successor->id()])) {
return false;
}
}
}
// Add successor phis.
if (mblock->successorWithPhis()) {
LBlock* phiSuccessor = mblock->successorWithPhis()->lir();
for (
unsigned int j = 0; j < phiSuccessor->numPhis(); j++) {
LPhi* phi = phiSuccessor->getPhi(j);
LAllocation* use = phi->getOperand(mblock->positionInPhiSuccessor());
uint32_t reg = use->toUse()->virtualRegister();
if (!live.insert(reg)) {
return false;
}
vreg(use).setUsedByPhi();
}
}
// Registers are assumed alive for the entire block, a define shortens
// the range to the point of definition.
for (VirtualRegBitSet::Iterator liveRegId(live); liveRegId; ++liveRegId) {
if (!vregs[*liveRegId].addInitialRange(alloc(), entryOf(block),
exitOf(block).next())) {
return false;
}
}
// Shorten the front end of ranges for live variables to their point of
// definition, if found.
for (LInstructionReverseIterator ins = block->rbegin();
ins != block->rend(); ins++) {
// Calls may clobber registers, so force a spill and reload around the
// callsite.
if (ins->isCall()) {
for (AnyRegisterIterator iter(allRegisters_.asLiveSet()); iter.more();
++iter) {
bool found =
false;
for (size_t i = 0; i < ins->numDefs(); i++) {
if (ins->getDef(i)->isFixed() &&
ins->getDef(i)->output()->aliases(LAllocation(*iter))) {
found =
true;
break;
}
}
// If this register doesn't have an explicit def above, mark
// it as clobbered by the call unless it is actually
// call-preserved.
if (!found && !ins->isCallPreserved(*iter)) {
if (!addInitialFixedRange(*iter, outputOf(*ins),
outputOf(*ins).next())) {
return false;
}
}
}
// Add the call position before the previous one to ensure the vector is
// sorted.
MOZ_ASSERT(prevCallPositionIndex > 0);
MOZ_ASSERT_IF(prevCallPositionIndex < callPositions.length(),
outputOf(*ins) < callPositions[prevCallPositionIndex]);
prevCallPositionIndex--;
callPositions[prevCallPositionIndex] = outputOf(*ins);
}
for (size_t i = 0; i < ins->numDefs(); i++) {
LDefinition* def = ins->getDef(i);
if (def->isBogusTemp()) {
continue;
}
CodePosition from = outputOf(*ins);
if (def->policy() == LDefinition::MUST_REUSE_INPUT) {
// MUST_REUSE_INPUT is implemented by allocating an output
// register and moving the input to it. Register hints are
// used to avoid unnecessary moves. We give the input an
// LUse::ANY policy to avoid allocating a register for the
// input.
LUse* inputUse = ins->getOperand(def->getReusedInput())->toUse();
MOZ_ASSERT(inputUse->policy() == LUse::
REGISTER);
MOZ_ASSERT(inputUse->usedAtStart());
*inputUse = LUse(inputUse->virtualRegister(), LUse::ANY,
/* usedAtStart = */ true);
}
if (!vreg(def).addInitialRange(alloc(), from, from.next())) {
return false;
}
vreg(def).setInitialDefinition(from);
live.remove(def->virtualRegister());
}
for (size_t i = 0; i < ins->numTemps(); i++) {
LDefinition* temp = ins->getTemp(i);
if (temp->isBogusTemp()) {
continue;
}
// Normally temps are considered to cover both the input
// and output of the associated instruction. In some cases
// though we want to use a fixed register as both an input
// and clobbered register in the instruction, so watch for
// this and shorten the temp to cover only the output.
CodePosition from = inputOf(*ins);
if (temp->policy() == LDefinition::FIXED) {
AnyRegister reg = temp->output()->toRegister();
for (LInstruction::InputIterator alloc(**ins); alloc.more();
alloc.next()) {
if (alloc->isUse()) {
LUse* use = alloc->toUse();
if (use->isFixedRegister()) {
if (GetFixedRegister(vreg(use).def(), use) == reg) {
from = outputOf(*ins);
}
}
}
}
}
// * For non-call instructions, temps cover both the input and output,
// so temps never alias uses (even at-start uses) or defs.
// * For call instructions, temps only cover the input (the output is
// used for the force-spill ranges added above). This means temps
// still don't alias uses but they can alias the (fixed) defs. For now
// we conservatively require temps to have a fixed register for call
// instructions to prevent a footgun.
MOZ_ASSERT_IF(ins->isCall(), temp->policy() == LDefinition::FIXED);
CodePosition to =
ins->isCall() ? outputOf(*ins) : outputOf(*ins).next();
if (!vreg(temp).addInitialRange(alloc(), from, to)) {
return false;
}
vreg(temp).setInitialDefinition(from);
}
DebugOnly<
bool> hasUseRegister =
false;
DebugOnly<
bool> hasUseRegisterAtStart =
false;
for (LInstruction::InputIterator inputAlloc(**ins); inputAlloc.more();
inputAlloc.next()) {
if (inputAlloc->isUse()) {
LUse* use = inputAlloc->toUse();
// Call uses should always be at-start, since calls use all
// registers.
MOZ_ASSERT_IF(ins->isCall() && !inputAlloc.isSnapshotInput(),
use->usedAtStart());
#ifdef DEBUG
// If there are both useRegisterAtStart(x) and useRegister(y)
// uses, we may assign the same register to both operands
// (bug 772830). Don't allow this for now.
if (use->policy() == LUse::
REGISTER) {
if (use->usedAtStart()) {
if (!IsInputReused(*ins, use)) {
hasUseRegisterAtStart =
true;
}
}
else {
hasUseRegister =
true;
}
}
MOZ_ASSERT(!(hasUseRegister && hasUseRegisterAtStart));
#endif
// Don't treat RECOVERED_INPUT uses as keeping the vreg alive.
if (use->policy() == LUse::RECOVERED_INPUT) {
continue;
}
CodePosition to = use->usedAtStart() ? inputOf(*ins) : outputOf(*ins);
if (use->isFixedRegister()) {
LAllocation reg(AnyRegister::FromCode(use->registerCode()));
for (size_t i = 0; i < ins->numDefs(); i++) {
LDefinition* def = ins->getDef(i);
if (def->policy() == LDefinition::FIXED &&
*def->output() == reg) {
to = inputOf(*ins);
}
}
}
if (!vreg(use).addInitialRange(alloc(), entryOf(block), to.next())) {
return false;
}
UsePosition* usePosition =
new (alloc().fallible()) UsePosition(use, to);
if (!usePosition) {
return false;
}
vreg(use).addInitialUse(usePosition);
if (!live.insert(use->virtualRegister())) {
return false;
}
}
}
}
// Phis have simultaneous assignment semantics at block begin, so at
// the beginning of the block we can be sure that liveIn does not
// contain any phi outputs.
for (
unsigned int i = 0; i < block->numPhis(); i++) {
LDefinition* def = block->getPhi(i)->getDef(0);
if (live.contains(def->virtualRegister())) {
live.remove(def->virtualRegister());
}
else {
// This is a dead phi, so add a dummy range over all phis. This
// can go away if we have an earlier dead code elimination pass.
CodePosition entryPos = entryOf(block);
if (!vreg(def).addInitialRange(alloc(), entryPos, entryPos.next())) {
return false;
}
}
}
if (mblock->isLoopHeader()) {
// MakeLoopsContiguous ensures blocks of a loop are contiguous, so add a
// single live range spanning the loop. Because we require liveIn in a
// later pass for resolution, we also have to fix that up for all loop
// blocks.
MBasicBlock* backedge = mblock->backedge();
// Add a range for this entire loop
CodePosition from = entryOf(mblock->lir());
CodePosition to = exitOf(backedge->lir()).next();
for (VirtualRegBitSet::Iterator liveRegId(live); liveRegId; ++liveRegId) {
if (!vregs[*liveRegId].addInitialRange(alloc(), from, to)) {
return false;
}
}
if (mblock != backedge) {
// Start at the block after |mblock|.
MOZ_ASSERT(graph.getBlock(i - 1) == mblock->lir());
size_t j = i;
while (
true) {
MBasicBlock* loopBlock = graph.getBlock(j)->mir();
// Fix up the liveIn set.
if (!liveIn[loopBlock->id()].insertAll(live)) {
return false;
}
if (loopBlock == backedge) {
break;
}
j++;
}
}
}
MOZ_ASSERT_IF(!mblock->numPredecessors(), live.empty());
}
MOZ_RELEASE_ASSERT(prevCallPositionIndex == 0,
"Must have initialized all call positions");
JitSpew(JitSpew_RegAlloc,
"Completed liveness analysis");
return true;
}
Maybe<size_t> BacktrackingAllocator::lookupFirstCallPositionInRange(
CodePosition from, CodePosition to) {
MOZ_ASSERT(from < to);
// Use binary search to find the first call position in range [from, to).
size_t len = callPositions.length();
size_t index;
mozilla::BinarySearch(callPositions, 0, len, from, &index);
MOZ_ASSERT(index <= len);
if (index == len) {
// Last call position in the vector is before this range.
MOZ_ASSERT_IF(len > 0, callPositions.back() < from);
return {};
}
if (callPositions[index] >= to) {
// The next call position comes after this range.
return {};
}
MOZ_ASSERT(callPositions[index] >= from && callPositions[index] < to);
return mozilla::Some(index);
}
///////////////////////////////////////////////////////////////////////////////
// //
// Merging and queueing of LiveRange groups //
// //
///////////////////////////////////////////////////////////////////////////////
// Helper for ::tryMergeBundles
static bool IsArgumentSlotDefinition(LDefinition* def) {
return def->policy() == LDefinition::FIXED && def->output()->isArgument();
}
// Helper for ::tryMergeBundles
static bool IsThisSlotDefinition(LDefinition* def) {
return IsArgumentSlotDefinition(def) &&
def->output()->toArgument()->index() <
THIS_FRAME_ARGSLOT +
sizeof(Value);
}
// Helper for ::tryMergeBundles
static bool HasStackPolicy(LDefinition* def) {
return def->policy() == LDefinition::STACK;
}
// Helper for ::tryMergeBundles
static bool CanMergeTypesInBundle(LDefinition::Type a, LDefinition::Type b) {
// Fast path for the common case.
if (a == b) {
return true;
}
// Only merge if the sizes match, so that we don't get confused about the
// width of spill slots.
return StackSlotAllocator::width(a) == StackSlotAllocator::width(b);
}
// Helper for ::tryMergeReusedRegister
bool BacktrackingAllocator::tryMergeBundles(LiveBundle* bundle0,
LiveBundle* bundle1) {
// See if bundle0 and bundle1 can be merged together.
if (bundle0 == bundle1) {
return true;
}
// Get a representative virtual register from each bundle.
VirtualRegister& reg0 = bundle0->firstRange()->vreg();
VirtualRegister& reg1 = bundle1->firstRange()->vreg();
MOZ_ASSERT(CanMergeTypesInBundle(reg0.type(), reg1.type()));
MOZ_ASSERT(reg0.isCompatible(reg1));
// Registers which might spill to the frame's |this| slot can only be
// grouped with other such registers. The frame's |this| slot must always
// hold the |this| value, as required by JitFrame tracing and by the Ion
// constructor calling convention.
if (IsThisSlotDefinition(reg0.def()) || IsThisSlotDefinition(reg1.def())) {
if (*reg0.def()->output() != *reg1.def()->output()) {
return true;
}
}
// Registers which might spill to the frame's argument slots can only be
// grouped with other such registers if the frame might access those
// arguments through a lazy arguments object or rest parameter.
if (IsArgumentSlotDefinition(reg0.def()) ||
IsArgumentSlotDefinition(reg1.def())) {
#ifdef JS_PUNBOX64
bool canSpillToArgSlots =
!graph.mir().entryBlock()->info().mayReadFrameArgsDirectly();
#else
bool canSpillToArgSlots =
false;
#endif
if (!canSpillToArgSlots) {
if (*reg0.def()->output() != *reg1.def()->output()) {
return true;
}
}
}
// When we make a call to a WebAssembly function that returns multiple
// results, some of those results can go on the stack. The callee is passed a
// pointer to this stack area, which is represented as having policy
// LDefinition::STACK (with type LDefinition::STACKRESULTS). Individual
// results alias parts of the stack area with a value-appropriate type, but
// policy LDefinition::STACK. This aliasing between allocations makes it
// unsound to merge anything with a LDefinition::STACK policy.
if (HasStackPolicy(reg0.def()) || HasStackPolicy(reg1.def())) {
return true;
}
// Limit the number of times we compare ranges if there are many ranges in
// one of the bundles, to avoid quadratic behavior.
static const size_t MAX_RANGES = 200;
// Make sure that ranges in the bundles do not overlap.
LiveBundle::RangeIterator iter0 = bundle0->rangesBegin(),
iter1 = bundle1->rangesBegin();
size_t count = 0;
while (iter0 && iter1) {
if (++count >= MAX_RANGES) {
return true;
}
LiveRange* range0 = *iter0;
LiveRange* range1 = *iter1;
if (range0->from() >= range1->to()) {
iter1++;
}
else if (range1->from() >= range0->to()) {
iter0++;
}
else {
return true;
}
}
// Move all ranges from bundle1 into bundle0. Use a fast path for the case
// where we can add all ranges from bundle1 at the end of bundle0.
if (SortBefore(bundle0->lastRange(), bundle1->firstRange())) {
while (LiveRange* range = bundle1->popFirstRange()) {
bundle0->addRangeAtEnd(range);
}
}
else {
LiveRange* prevRange = nullptr;
while (LiveRange* range = bundle1->popFirstRange()) {
bundle0->addRange(range, prevRange);
prevRange = range;
}
}
return true;
}
// Helper for ::mergeAndQueueRegisters
void BacktrackingAllocator::allocateStackDefinition(VirtualRegister& reg) {
LInstruction* ins = reg.ins()->toInstruction();
if (reg.def()->type() == LDefinition::STACKRESULTS) {
LStackArea alloc(ins->toInstruction());
stackSlotAllocator.allocateStackArea(&alloc);
reg.def()->setOutput(alloc);
}
else {
// Because the definitions are visited in order, the area has been allocated
// before we reach this result, so we know the operand is an LStackArea.
const LUse* use = ins->getOperand(0)->toUse();
VirtualRegister& area = vregs[use->virtualRegister()];
const LStackArea* areaAlloc = area.def()->output()->toStackArea();
reg.def()->setOutput(areaAlloc->resultAlloc(ins, reg.def()));
}
}
// Helper for ::mergeAndQueueRegisters
bool BacktrackingAllocator::tryMergeReusedRegister(VirtualRegister& def,
VirtualRegister& input) {
// def is a vreg which reuses input for its output physical register. Try
// to merge ranges for def with those of input if possible, as avoiding
// copies before def's instruction is crucial for generated code quality
// (MUST_REUSE_INPUT is used for all arithmetic on x86/x64).
// Don't try to merge if the definition starts at the input point of the
// instruction.
if (def.firstRange()->from() == inputOf(def.ins())) {
MOZ_ASSERT(def.isTemp());
def.setMustCopyInput();
return true;
}
MOZ_ASSERT(def.firstRange()->from() == outputOf(def.ins()));
if (!CanMergeTypesInBundle(def.type(), input.type())) {
def.setMustCopyInput();
return true;
}
LiveRange* inputRange = input.rangeFor(outputOf(def.ins()));
if (!inputRange) {
// The input is not live after the instruction, either in a safepoint
// for the instruction or in subsequent code. The input and output
// can thus be in the same group.
return tryMergeBundles(def.firstBundle(), input.firstBundle());
}
// Avoid merging in very large live ranges as merging has non-linear
// complexity. The cutoff value is hard to gauge. 1M was chosen because it
// is "large" and yet usefully caps compile time on AutoCad-for-the-web to
// something reasonable on a 2017-era desktop system. See bug 1728781.
//
// Subsequently -- bug 1897427 comment 9 -- even this threshold seems too
// high, resulting in delays of around 10 seconds. Setting it to 500000
// "fixes" that one, but it seems prudent to reduce it to 250000 so as to
// avoid having to change it yet again at some future point.
const uint32_t RANGE_SIZE_CUTOFF = 250000;
if (inputRange->to() - inputRange->from() > RANGE_SIZE_CUTOFF) {
def.setMustCopyInput();
return true;
}
// The input is live afterwards, either in future instructions or in a
// safepoint for the reusing instruction. This is impossible to satisfy
// without copying the input.
//
// It may or may not be better to split the input into two bundles at the
// point of the definition, which may permit merging. One case where it is
// definitely better to split is if the input never has any register uses
// after the instruction. Handle this splitting eagerly.
LBlock* block = def.ins()->block();
// The input's lifetime must end within the same block as the definition,
// otherwise it could live on in phis elsewhere.
if (inputRange != input.lastRange() || inputRange->to() > exitOf(block)) {
def.setMustCopyInput();
return true;
}
// If we already split the input for some other register, don't make a
// third bundle.
if (inputRange->bundle() != input.firstRange()->bundle()) {
def.setMustCopyInput();
return true;
}
// If the input will start out in memory then adding a separate bundle for
// memory uses after the def won't help.
if (input.def()->isFixed() && !input.def()->output()->isRegister()) {
def.setMustCopyInput();
return true;
}
// The input cannot have register or reused uses after the definition.
for (UsePositionIterator iter = inputRange->usesBegin(); iter; iter++) {
if (iter->pos <= inputOf(def.ins())) {
continue;
}
LUse* use = iter->use();
if (FindReusingDefOrTemp(insData[iter->pos], use)) {
def.setMustCopyInput();
return true;
}
if (iter->usePolicy() != LUse::ANY &&
iter->usePolicy() != LUse::KEEPALIVE) {
def.setMustCopyInput();
return true;
}
}
LiveRange* preRange = LiveRange::FallibleNew(
alloc(), &input, inputRange->from(), outputOf(def.ins()));
if (!preRange) {
return false;
}
// The new range starts at reg's input position, which means it overlaps
// with the old range at one position. This is what we want, because we
// need to copy the input before the instruction.
LiveRange* postRange = LiveRange::FallibleNew(
alloc(), &input, inputOf(def.ins()), inputRange->to());
if (!postRange) {
return false;
}
inputRange->tryToMoveDefAndUsesInto(preRange);
inputRange->tryToMoveDefAndUsesInto(postRange);
MOZ_ASSERT(!inputRange->hasUses());
JitSpewIfEnabled(JitSpew_RegAlloc,
" splitting reused input at %u to try to help grouping",
inputOf(def.ins()).bits());
LiveBundle* firstBundle = inputRange->bundle();
// Note: this inserts an item at the beginning of the ranges vector. That's
// okay in this case because the checks above ensure this happens at most once
// for the |input| virtual register.
if (!input.replaceLastRangeLinear(inputRange, preRange, postRange)) {
return false;
}
firstBundle->removeRange(inputRange);
firstBundle->addRange(preRange);
// The new range goes in a separate bundle, where it will be spilled during
// allocation.
LiveBundle* secondBundle =
LiveBundle::FallibleNew(alloc(), nullptr, nullptr, getNextBundleId());
if (!secondBundle) {
return false;
}
secondBundle->addRangeAtEnd(postRange);
return tryMergeBundles(def.firstBundle(), input.firstBundle());
}
bool BacktrackingAllocator::mergeAndQueueRegisters() {
MOZ_ASSERT(!vregs[0u].hasRanges());
// Create a bundle for each register containing all its ranges.
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (!reg.hasRanges()) {
continue;
}
LiveBundle* bundle =
LiveBundle::FallibleNew(alloc(), nullptr, nullptr, getNextBundleId());
if (!bundle) {
return false;
}
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
bundle->addRangeAtEnd(*iter);
}
}
// If there is an OSR block, merge parameters in that block with the
// corresponding parameters in the initial block.
if (MBasicBlock* osr = graph.mir().osrBlock()) {
size_t original = 1;
for (LInstructionIterator iter = osr->lir()->begin();
iter != osr->lir()->end(); iter++) {
if (iter->isParameter()) {
for (size_t i = 0; i < iter->numDefs(); i++) {
DebugOnly<
bool> found =
false;
VirtualRegister& paramVreg = vreg(iter->getDef(i));
for (; original < paramVreg.vreg(); original++) {
VirtualRegister& originalVreg = vregs[original];
if (*originalVreg.def()->output() == *iter->getDef(i)->output()) {
MOZ_ASSERT(originalVreg.ins()->isParameter());
if (!tryMergeBundles(originalVreg.firstBundle(),
paramVreg.firstBundle())) {
return false;
}
found =
true;
break;
}
}
MOZ_ASSERT(found);
}
}
}
}
// Try to merge registers with their reused inputs.
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (!reg.hasRanges()) {
continue;
}
if (reg.def()->policy() == LDefinition::MUST_REUSE_INPUT) {
LUse* use = reg.ins()
->toInstruction()
->getOperand(reg.def()->getReusedInput())
->toUse();
if (!tryMergeReusedRegister(reg, vreg(use))) {
return false;
}
}
}
// Try to merge phis with their inputs.
for (size_t i = 0; i < graph.numBlocks(); i++) {
LBlock* block = graph.getBlock(i);
for (size_t j = 0; j < block->numPhis(); j++) {
LPhi* phi = block->getPhi(j);
VirtualRegister& outputVreg = vreg(phi->getDef(0));
for (size_t k = 0, kend = phi->numOperands(); k < kend; k++) {
VirtualRegister& inputVreg = vreg(phi->getOperand(k)->toUse());
if (!tryMergeBundles(inputVreg.firstBundle(),
outputVreg.firstBundle())) {
return false;
}
}
}
}
// Add all bundles to the allocation queue, and create spill sets for them.
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
// Eagerly allocate stack result areas and their component stack results.
if (reg.def() && reg.def()->policy() == LDefinition::STACK) {
allocateStackDefinition(reg);
}
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
LiveRange* range = *iter;
LiveBundle* bundle = range->bundle();
if (range == bundle->firstRange()) {
if (!alloc().ensureBallast()) {
return false;
}
SpillSet* spill = SpillSet::
New(alloc());
if (!spill) {
return false;
}
bundle->setSpillSet(spill);
size_t priority = computePriority(bundle);
if (!allocationQueue.insert(QueueItem(bundle, priority))) {
return false;
}
}
}
}
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Code for the splitting/spilling subsystem begins here. //
// //
// The code that follows is structured in the following sequence: //
// //
// (1) Routines that are helpers for ::splitAt. //
// (2) ::splitAt itself, which implements splitting decisions. //
// (3) heuristic routines (eg ::splitAcrossCalls), which decide where //
// splits should be made. They then call ::splitAt to perform the //
// chosen split. //
// (4) The top level driver, ::chooseBundleSplit. //
// //
// There are further comments on ::splitAt and ::chooseBundleSplit below. //
// //
///////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////
// //
// Implementation of splitting decisions, but not the making of those //
// decisions: various helper functions //
// //
///////////////////////////////////////////////////////////////////////////////
bool BacktrackingAllocator::updateVirtualRegisterListsThenRequeueBundles(
LiveBundle* bundle,
const LiveBundleVector& newBundles) {
#ifdef DEBUG
if (newBundles.length() == 1) {
LiveBundle* newBundle = newBundles[0];
if (newBundle->numRanges() == bundle->numRanges() &&
computePriority(newBundle) == computePriority(bundle)) {
bool different =
false;
LiveBundle::RangeIterator oldRanges = bundle->rangesBegin();
LiveBundle::RangeIterator newRanges = newBundle->rangesBegin();
while (oldRanges) {
LiveRange* oldRange = *oldRanges;
LiveRange* newRange = *newRanges;
if (oldRange->from() != newRange->from() ||
oldRange->to() != newRange->to()) {
different =
true;
break;
}
oldRanges++;
newRanges++;
}
// This is likely to trigger an infinite loop in register allocation. This
// can be the result of invalid register constraints, making regalloc
// impossible; consider relaxing those.
MOZ_ASSERT(different,
"Split results in the same bundle with the same priority");
}
}
#endif
if (JitSpewEnabled(JitSpew_RegAlloc)) {
JitSpew(JitSpew_RegAlloc,
" .. into:");
for (size_t i = 0; i < newBundles.length(); i++) {
JitSpew(JitSpew_RegAlloc,
" %s", newBundles[i]->toString().get());
}
}
// Remove all ranges in the old bundle from their register's list.
bundle->removeAllRangesFromVirtualRegisters();
// Add all ranges in the new bundles to their register's list.
for (size_t i = 0; i < newBundles.length(); i++) {
LiveBundle* newBundle = newBundles[i];
for (LiveBundle::RangeIterator iter = newBundle->rangesBegin(); iter;
iter++) {
LiveRange* range = *iter;
if (!range->vreg().addRange(range)) {
return false;
}
}
}
// Queue the new bundles for register assignment.
for (size_t i = 0; i < newBundles.length(); i++) {
LiveBundle* newBundle = newBundles[i];
size_t priority = computePriority(newBundle);
if (!allocationQueue.insert(QueueItem(newBundle, priority))) {
return false;
}
}
return true;
}
// Helper for ::splitAt
// When splitting a bundle according to a list of split positions, return
// whether a use or range at |pos| should use a different bundle than the last
// position this was called for.
static bool UseNewBundle(
const SplitPositionVector& splitPositions,
CodePosition pos, size_t* activeSplitPosition) {
if (splitPositions.empty()) {
// When the split positions are empty we are splitting at all uses.
return true;
}
if (*activeSplitPosition == splitPositions.length()) {
// We've advanced past all split positions.
return false;
}
if (splitPositions[*activeSplitPosition] > pos) {
// We haven't gotten to the next split position yet.
return false;
}
// We've advanced past the next split position, find the next one which we
// should split at.
while (*activeSplitPosition < splitPositions.length() &&
splitPositions[*activeSplitPosition] <= pos) {
(*activeSplitPosition)++;
}
return true;
}
// Helper for ::splitAt
static bool HasPrecedingRangeSharingVreg(LiveBundle* bundle, LiveRange* range) {
MOZ_ASSERT(range->bundle() == bundle);
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* prevRange = *iter;
if (prevRange == range) {
return false;
}
if (&prevRange->vreg() == &range->vreg()) {
return true;
}
}
MOZ_CRASH();
}
// Helper for ::splitAt
static bool HasFollowingRangeSharingVreg(LiveBundle* bundle, LiveRange* range) {
MOZ_ASSERT(range->bundle() == bundle);
LiveBundle::RangeIterator iter = bundle->rangesBegin(range);
MOZ_ASSERT(*iter == range);
iter++;
for (; iter; iter++) {
LiveRange* nextRange = *iter;
if (&nextRange->vreg() == &range->vreg()) {
return true;
}
}
return false;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Implementation of splitting decisions, but not the making of those //
// decisions: //
// ::splitAt //
// //
///////////////////////////////////////////////////////////////////////////////
// ::splitAt
// ---------
// It would be nice to be able to interpret ::splitAt as simply performing
// whatever split the heuristic routines decide on. Unfortunately it
// tries to "improve" on the initial locations, which as
// https://bugzilla.mozilla.org/show_bug.cgi?id=1758274#c17 shows, often
// leads to excessive spilling. So there is no clean distinction between
// policy (where to split, computed by the heuristic routines) and
// implementation (done by ::splitAt).
//
// ::splitAt -- creation of spill parent bundles
// ---------------------------------------------
// To understand what ::splitAt does, we must refer back to Section 1's
// description of LiveBundle::spillParent_.
//
// Initially (as created by Phase 1), all bundles have `spillParent_` being
// NULL. If ::splitAt is asked to split such a bundle, it will first create a
// "spill bundle" or "spill parent" bundle. This is a copy of the original,
// with two changes:
//
// * all register uses have been removed, so that only stack-compatible uses
// remain.
//
// * for all LiveRanges in the bundle that define a register, the start point
// of the range is moved one CodePosition forwards, thusly:
//
// from = minimalDefEnd(insData[from]).next();
//
// The reason for the latter relates to the idea described in Section 1, that
// all LiveRanges for any given VirtualRegister must form a tree rooted at the
// defining LiveRange. If the spill-bundle definition range start points are
// the same as those in the original bundle, then we will end up with two roots
// for the tree, and it is then unclear which one should supply "the value".
//
// Putting the spill-bundle start point one CodePosition further along causes
// the range containing the register def (after splitting) to still be the
// defining point. ::createMoveGroupsFromLiveRangeTransitions will observe the
// equivalent spill-bundle range starting one point later and add a MoveGroup
// to move the value into it. Since the spill bundle is intended to be stack
// resident, the effect is to force creation of the MoveGroup that will
// actually spill this value onto the stack.
//
// If the bundle provided to ::splitAt already has a spill parent, then
// ::splitAt doesn't create a new spill parent. This situation will happen
// when the bundle to be split was itself created by splitting. The effect is
// that *all* bundles created from an "original bundle" share the same spill
// parent, and hence they will share the same spill slot, which guarantees that
// all the spilled fragments of a VirtualRegister share the same spill slot,
// which means we'll never have to move a VirtualRegister between different
// spill slots during its lifetime.
//
// ::splitAt -- creation of other bundles
// --------------------------------------
// With the spill parent bundle question out of the way, ::splitAt then goes on
// to create the remaining new bundles, using near-incomprehensible logic
// steered by `UseNewBundle`.
//
// This supposedly splits the bundle at the positions given by the
// `SplitPositionVector` parameter to ::splitAt, putting them in a temporary
// vector `newBundles`. Whether it really splits at the requested positions or
// not is hard to say; more important is what happens next.
//
// ::splitAt -- "improvement" ("filtering") of the split bundles
// -------------------------------------------------------------
// ::splitAt now tries to reduce the length of the LiveRanges that make up the
// new bundles (not including the "spill parent"). I assume this is to remove
// sections of the bundles that carry no useful value (eg, extending after the
// last using a range), thereby removing the demand for registers in those
// parts. This does however mean that ::splitAt is no longer really splitting
// where the heuristic routines wanted, and that can lead to a big increase in
// spilling in loops, as
// https://bugzilla.mozilla.org/show_bug.cgi?id=1758274#c17 describes.
//
// ::splitAt -- meaning of the incoming `SplitPositionVector`
// ----------------------------------------------------------
// ::splitAt has one last mystery which is important to document. The split
// positions are specified as CodePositions, but this leads to ambiguity
// because, in a sequence of N (LIR) instructions, there are 2N valid
// CodePositions. For example:
//
// 6-7 WasmLoadTls [def v2<o>] [use v1:R]
// 8-9 WasmNullConstant [def v3<o>]
//
// Consider splitting the range for `v2`, which starts at CodePosition 7.
// What's the difference between saying "split it at 7" and "split it at 8" ?
// Not much really, since in both cases what we intend is for the range to be
// split in between the two instructions.
//
// Hence I believe the semantics is:
//
// * splitting at an even numbered CodePosition (eg, 8), which is an input-side
// position, means "split before the instruction containing this position".
//
// * splitting at an odd numbered CodePositin (eg, 7), which is an output-side
// position, means "split after the instruction containing this position".
//
// Hence in the example, we could specify either 7 or 8 to mean the same
// placement of the split. Well, almost true, but actually:
//
// (SPEC) specifying 8 means
//
// "split between these two insns, and any resulting MoveGroup goes in the
// list to be emitted before the start of the second insn"
//
// (SPEC) specifying 7 means
//
// "split between these two insns, and any resulting MoveGroup goes in the
// list to be emitted after the end of the first insn"
//
// In most cases we don't care on which "side of the valley" the MoveGroup ends
// up, in which case we can use either convention.
//
// (SPEC) I believe these semantics are implied by the logic in
// ::createMoveGroupsFromLiveRangeTransitions. They are certainly not
// documented anywhere in the code.
bool BacktrackingAllocator::splitAt(LiveBundle* bundle,
const SplitPositionVector& splitPositions) {
// Split the bundle at the given split points. Register uses which have no
// intervening split points are consolidated into the same bundle. If the
// list of split points is empty, then all register uses are placed in
// minimal bundles.
// splitPositions should be sorted.
for (size_t i = 1; i < splitPositions.length(); ++i) {
MOZ_ASSERT(splitPositions[i - 1] < splitPositions[i]);
}
// We don't need to create a new spill bundle if there already is one.
bool spillBundleIsNew =
false;
LiveBundle* spillBundle = bundle->spillParent();
if (!spillBundle) {
spillBundle = LiveBundle::FallibleNew(alloc(), bundle->spillSet(), nullptr,
getNextBundleId());
if (!spillBundle) {
return false;
}
spillBundleIsNew =
true;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
CodePosition from = range->from();
if (isRegisterDefinition(range)) {
from = minimalDefEnd(insData[from]).next();
}
if (from < range->to()) {
if (!spillBundle->addRangeAtEnd(alloc(), &range->vreg(), from,
range->to())) {
return false;
}
if (range->hasDefinition() && !isRegisterDefinition(range)) {
spillBundle->lastRange()->setHasDefinition();
}
}
}
}
LiveBundleVector newBundles;
// The bundle which ranges are currently being added to.
LiveBundle* activeBundle = LiveBundle::FallibleNew(
alloc(), bundle->spillSet(), spillBundle, getNextBundleId());
if (!activeBundle || !newBundles.append(activeBundle)) {
return false;
}
// State for use by UseNewBundle.
size_t activeSplitPosition = 0;
// Make new bundles according to the split positions, and distribute ranges
// and uses to them.
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (UseNewBundle(splitPositions, range->from(), &activeSplitPosition)) {
activeBundle = LiveBundle::FallibleNew(alloc(), bundle->spillSet(),
spillBundle, getNextBundleId());
if (!activeBundle || !newBundles.append(activeBundle)) {
return false;
}
}
LiveRange* activeRange = LiveRange::FallibleNew(alloc(), &range->vreg(),
range->from(), range->to());
if (!activeRange) {
return false;
}
activeBundle->addRangeAtEnd(activeRange);
if (isRegisterDefinition(range)) {
activeRange->setHasDefinition();
}
while (range->hasUses()) {
UsePosition* use = range->popUse();
LNode* ins = insData[use->pos];
// Any uses of a register that appear before its definition has
// finished must be associated with the range for that definition.
if (isRegisterDefinition(range) &&
use->pos <= minimalDefEnd(insData[range->from()])) {
activeRange->addUse(use);
}
else if (isRegisterUse(use, ins)) {
// Place this register use into a different bundle from the
// last one if there are any split points between the two uses.
// UseNewBundle always returns true if we are splitting at all
// register uses, but we can still reuse the last range and
// bundle if they have uses at the same position, except when
// either use is fixed (the two uses might require incompatible
// registers.)
if (UseNewBundle(splitPositions, use->pos, &activeSplitPosition) &&
(!activeRange->hasUses() ||
activeRange->usesBegin()->pos != use->pos ||
activeRange->usesBegin()->usePolicy() == LUse::FIXED ||
use->usePolicy() == LUse::FIXED)) {
activeBundle = LiveBundle::FallibleNew(
alloc(), bundle->spillSet(), spillBundle, getNextBundleId());
if (!activeBundle || !newBundles.append(activeBundle)) {
return false;
}
activeRange = LiveRange::FallibleNew(alloc(), &range->vreg(),
range->from(), range->to());
if (!activeRange) {
return false;
}
activeBundle->addRangeAtEnd(activeRange);
}
activeRange->addUse(use);
}
else {
MOZ_ASSERT(spillBundleIsNew);
spillBundle->rangeFor(use->pos)->addUse(use);
}
}
}
LiveBundleVector filteredBundles;
// Trim the ends of ranges in each new bundle when there are no other
// earlier or later ranges in the same bundle with the same vreg.
for (size_t i = 0; i < newBundles.length(); i++) {
LiveBundle* bundle = newBundles[i];
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter;) {
LiveRange* range = *iter;
if (!range->hasDefinition()) {
if (!HasPrecedingRangeSharingVreg(bundle, range)) {
if (range->hasUses()) {
UsePosition* use = *range->usesBegin();
range->setFrom(inputOf(insData[use->pos]));
}
else {
bundle->removeRangeAndIncrementIterator(iter);
continue;
}
}
}
if (!HasFollowingRangeSharingVreg(bundle, range)) {
if (range->hasUses()) {
UsePosition* use = range->lastUse();
range->setTo(use->pos.next());
}
else if (range->hasDefinition()) {
range->setTo(minimalDefEnd(insData[range->from()]).next());
}
else {
bundle->removeRangeAndIncrementIterator(iter);
continue;
}
}
iter++;
}
if (bundle->hasRanges() && !filteredBundles.append(bundle)) {
return false;
}
}
if (spillBundleIsNew && !filteredBundles.append(spillBundle)) {
return false;
}
return updateVirtualRegisterListsThenRequeueBundles(bundle, filteredBundles);
}
///////////////////////////////////////////////////////////////////////////////
// //
// Creation of splitting decisions, but not their implementation: //
// ::splitAcrossCalls //
// ::trySplitAcrossHotcode //
// ::trySplitAfterLastRegisterUse //
// ::trySplitBeforeFirstRegisterUse //
// //
///////////////////////////////////////////////////////////////////////////////
bool BacktrackingAllocator::splitAcrossCalls(LiveBundle* bundle) {
// Split the bundle to separate register uses and non-register uses and
// allow the vreg to be spilled across its range.
// Locations of all calls in the bundle's range. This vector is sorted in
// ascending order because ranges in the bundle are sorted by start position
// and don't overlap
SplitPositionVector bundleCallPositions;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
// Append call positions in range [from+1, to) to bundleCallPositions.
// Calls at the beginning of the range are ignored because there is no
// splitting to do in this case.
CodePosition from = range->from().next();
if (from == range->to()) {
// Empty range can't contain any calls.
continue;
}
// Use binary search to find the first call position.
Maybe<size_t> index = lookupFirstCallPositionInRange(from, range->to());
if (!index.isSome()) {
// There are no calls inside this range.
continue;
}
// Find the index of the last call within this range.
size_t startIndex = *index;
size_t endIndex = startIndex;
while (endIndex < callPositions.length() - 1) {
if (callPositions[endIndex + 1] >= range->to()) {
break;
}
endIndex++;
}
MOZ_ASSERT(startIndex <= endIndex);
#ifdef DEBUG
auto inRange = [range](CodePosition pos) {
return range->covers(pos) && pos != range->from();
};
// Assert startIndex is the first call position in this range.
MOZ_ASSERT(inRange(callPositions[startIndex]));
MOZ_ASSERT_IF(startIndex > 0, !inRange(callPositions[startIndex - 1]));
// Assert endIndex is the last call position in this range.
MOZ_ASSERT(inRange(callPositions[endIndex]));
MOZ_ASSERT_IF(endIndex + 1 < callPositions.length(),
!inRange(callPositions[endIndex + 1]));
// Assert bundleCallPositions stays sorted.
MOZ_ASSERT_IF(!bundleCallPositions.empty(),
bundleCallPositions.back() < callPositions[startIndex]);
#endif
const CodePosition* start = &callPositions[startIndex];
size_t count = endIndex - startIndex + 1;
if (!bundleCallPositions.append(start, start + count)) {
return false;
}
}
MOZ_ASSERT(!bundleCallPositions.empty());
#ifdef JS_JITSPEW
JitSpewStart(JitSpew_RegAlloc,
" .. split across calls at ");
for (size_t i = 0; i < bundleCallPositions.length(); ++i) {
JitSpewCont(JitSpew_RegAlloc,
"%s%u", i != 0 ?
", " :
"",
bundleCallPositions[i].bits());
}
JitSpewFin(JitSpew_RegAlloc);
#endif
return splitAt(bundle, bundleCallPositions);
}
bool BacktrackingAllocator::trySplitAcrossHotcode(LiveBundle* bundle,
bool* success) {
// If this bundle has portions that are hot and portions that are cold,
// split it at the boundaries between hot and cold code.
LiveRange* hotRange = nullptr;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (hotcode.contains(range, &hotRange)) {
break;
}
}
// Don't split if there is no hot code in the bundle.
if (!hotRange) {
JitSpew(JitSpew_RegAlloc,
" .. bundle does not contain hot code");
return true;
}
// Don't split if there is no cold code in the bundle.
bool coldCode =
false;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (!hotRange->contains(range)) {
coldCode =
true;
break;
}
}
if (!coldCode) {
JitSpew(JitSpew_RegAlloc,
" .. bundle does not contain cold code");
return true;
}
JitSpewIfEnabled(JitSpew_RegAlloc,
" .. split across hot range %s",
hotRange->toString().get());
// Tweak the splitting method when compiling wasm code to look at actual
// uses within the hot/cold code. This heuristic is in place as the below
// mechanism regresses several asm.js tests. Hopefully this will be fixed
// soon and this special case removed. See bug 948838.
if (compilingWasm()) {
SplitPositionVector splitPositions;
if (!splitPositions.append(hotRange->from()) ||
!splitPositions.append(hotRange->to())) {
return false;
}
*success =
true;
return splitAt(bundle, splitPositions);
}
LiveBundle* hotBundle = LiveBundle::FallibleNew(
alloc(), bundle->spillSet(), bundle->spillParent(), getNextBundleId());
if (!hotBundle) {
return false;
}
LiveBundle* preBundle = nullptr;
LiveBundle* postBundle = nullptr;
LiveBundle* coldBundle = nullptr;
if (testbed) {
coldBundle = LiveBundle::FallibleNew(
alloc(), bundle->spillSet(), bundle->spillParent(), getNextBundleId());
if (!coldBundle) {
return false;
}
}
// Accumulate the ranges of hot and cold code in the bundle. Note that
// we are only comparing with the single hot range found, so the cold code
// may contain separate hot ranges.
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
LiveRange::Range hot, coldPre, coldPost;
range->intersect(hotRange, &coldPre, &hot, &coldPost);
if (!hot.empty()) {
if (!hotBundle->addRangeAndDistributeUses(alloc(), range, hot.from,
hot.to)) {
return false;
}
}
if (!coldPre.empty()) {
if (testbed) {
if (!coldBundle->addRangeAndDistributeUses(alloc(), range, coldPre.from,
coldPre.to)) {
return false;
}
}
else {
if (!preBundle) {
preBundle =
LiveBundle::FallibleNew(alloc(), bundle->spillSet(),
bundle->spillParent(), getNextBundleId());
if (!preBundle) {
return false;
}
}
if (!preBundle->addRangeAndDistributeUses(alloc(), range, coldPre.from,
coldPre.to)) {
return false;
}
}
}
if (!coldPost.empty()) {
if (testbed) {
if (!coldBundle->addRangeAndDistributeUses(
alloc(), range, coldPost.from, coldPost.to)) {
return false;
}
}
else {
if (!postBundle) {
postBundle =
LiveBundle::FallibleNew(alloc(), bundle->spillSet(),
bundle->spillParent(), getNextBundleId());
if (!postBundle) {
return false;
}
}
if (!postBundle->addRangeAndDistributeUses(
alloc(), range, coldPost.from, coldPost.to)) {
return false;
}
}
}
}
MOZ_ASSERT(hotBundle->numRanges() != 0);
LiveBundleVector newBundles;
if (!newBundles.append(hotBundle)) {
return false;
}
if (testbed) {
MOZ_ASSERT(coldBundle->numRanges() != 0);
if (!newBundles.append(coldBundle)) {
return false;
}
}
else {
MOZ_ASSERT(preBundle || postBundle);
if (preBundle && !newBundles.append(preBundle)) {
return false;
}
if (postBundle && !newBundles.append(postBundle)) {
return false;
}
}
*success =
true;
return updateVirtualRegisterListsThenRequeueBundles(bundle, newBundles);
}
bool BacktrackingAllocator::trySplitAfterLastRegisterUse(LiveBundle* bundle,
LiveBundle* conflict,
bool* success) {
// If this bundle's later uses do not require it to be in a register,
// split it after the last use which does require a register. If conflict
// is specified, only consider register uses before the conflict starts.
CodePosition lastRegisterFrom, lastRegisterTo, lastUse;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
// If the range defines a register, consider that a register use for
// our purposes here.
if (isRegisterDefinition(range)) {
CodePosition spillStart = minimalDefEnd(insData[range->from()]).next();
if (!conflict || spillStart < conflict->firstRange()->from()) {
lastUse = lastRegisterFrom = range->from();
lastRegisterTo = spillStart;
}
}
for (UsePositionIterator iter(range->usesBegin()); iter; iter++) {
LNode* ins = insData[iter->pos];
// Uses in the bundle should be sorted.
MOZ_ASSERT(iter->pos >= lastUse);
lastUse = inputOf(ins);
if (!conflict || outputOf(ins) < conflict->firstRange()->from()) {
if (isRegisterUse(*iter, ins,
/* considerCopy = */ true)) {
lastRegisterFrom = inputOf(ins);
lastRegisterTo = iter->pos.next();
}
}
}
}
// Can't trim non-register uses off the end by splitting.
if (!lastRegisterFrom.bits()) {
JitSpew(JitSpew_RegAlloc,
" .. bundle has no register uses");
return true;
}
if (lastUse < lastRegisterTo) {
JitSpew(JitSpew_RegAlloc,
" .. bundle's last use is a register use");
return true;
}
JitSpewIfEnabled(JitSpew_RegAlloc,
" .. split after last register use at %u",
lastRegisterTo.bits());
SplitPositionVector splitPositions;
if (!splitPositions.append(lastRegisterTo)) {
return false;
}
*success =
true;
return splitAt(bundle, splitPositions);
}
bool BacktrackingAllocator::trySplitBeforeFirstRegisterUse(LiveBundle* bundle,
LiveBundle* conflict,
bool* success) {
// If this bundle's earlier uses do not require it to be in a register,
// split it before the first use which does require a register. If conflict
// is specified, only consider register uses after the conflict ends.
if (isRegisterDefinition(bundle->firstRange())) {
JitSpew(JitSpew_RegAlloc,
" .. bundle is defined by a register");
return true;
}
if (!bundle->firstRange()->hasDefinition()) {
JitSpew(JitSpew_RegAlloc,
" .. bundle does not have definition");
return true;
}
CodePosition firstRegisterFrom;
CodePosition conflictEnd;
if (conflict) {
for (LiveBundle::RangeIterator iter = conflict->rangesBegin(); iter;
iter++) {
LiveRange* range = *iter;
if (range->to() > conflictEnd) {
conflictEnd = range->to();
}
}
}
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (!conflict || range->from() > conflictEnd) {
if (range->hasDefinition() && isRegisterDefinition(range)) {
firstRegisterFrom = range->from();
break;
}
}
for (UsePositionIterator iter(range->usesBegin()); iter; iter++) {
LNode* ins = insData[iter->pos];
if (!conflict || outputOf(ins) >= conflictEnd) {
if (isRegisterUse(*iter, ins,
/* considerCopy = */ true)) {
firstRegisterFrom = inputOf(ins);
break;
}
}
}
if (firstRegisterFrom.bits()) {
break;
}
}
if (!firstRegisterFrom.bits()) {
// Can't trim non-register uses off the beginning by splitting.
JitSpew(JitSpew_RegAlloc,
" bundle has no register uses");
return true;
}
JitSpewIfEnabled(JitSpew_RegAlloc,
" .. split before first register use at %u",
firstRegisterFrom.bits());
SplitPositionVector splitPositions;
if (!splitPositions.append(firstRegisterFrom)) {
return false;
}
*success =
true;
return splitAt(bundle, splitPositions);
}
///////////////////////////////////////////////////////////////////////////////
// //
// The top level driver for the splitting machinery //
// //
///////////////////////////////////////////////////////////////////////////////
// ::chooseBundleSplit
// -------------------
// If the first allocation loop (in ::go) can't allocate a bundle, it hands it
// off to ::chooseBundleSplit, which is the "entry point" of the bundle-split
// machinery. This tries four heuristics in turn, to see if any can split the
// bundle:
//
// * ::trySplitAcrossHotcode
// * ::splitAcrossCalls (in some cases)
// * ::trySplitBeforeFirstRegisterUse
// * ::trySplitAfterLastRegisterUse
//
// These routines have similar structure: they try to decide on one or more
// CodePositions at which to split the bundle, using whatever heuristics they
// have to hand. If suitable CodePosition(s) are found, they are put into a
// `SplitPositionVector`, and the bundle and the vector are handed off to
// ::splitAt, which performs the split (at least in theory) at the chosen
// positions. It also arranges for the new bundles to be added to
// ::allocationQueue.
//
// ::trySplitAcrossHotcode has a special case for JS -- it modifies the
// bundle(s) itself, rather than using ::splitAt.
//
// If none of the heuristic routines apply, then ::splitAt is called with an
// empty vector of split points. This is interpreted to mean "split at all
// register uses". When combined with how ::splitAt works, the effect is to
// spill the bundle.
bool BacktrackingAllocator::chooseBundleSplit(LiveBundle* bundle,
bool hasCall,
LiveBundle* conflict) {
bool success =
false;
JitSpewIfEnabled(JitSpew_RegAlloc,
" Splitting %s ..",
bundle->toString().get());
if (!trySplitAcrossHotcode(bundle, &success)) {
return false;
}
if (success) {
return true;
}
if (hasCall) {
return splitAcrossCalls(bundle);
}
if (!trySplitBeforeFirstRegisterUse(bundle, conflict, &success)) {
return false;
}
if (success) {
return true;
}
if (!trySplitAfterLastRegisterUse(bundle, conflict, &success)) {
return false;
}
if (success) {
return true;
}
// Split at all register uses.
SplitPositionVector emptyPositions;
return splitAt(bundle, emptyPositions);
}
///////////////////////////////////////////////////////////////////////////////
// //
// Bundle allocation //
// //
///////////////////////////////////////////////////////////////////////////////
static const size_t MAX_ATTEMPTS = 2;
bool BacktrackingAllocator::computeRequirement(LiveBundle* bundle,
Requirement* requirement,
Requirement* hint) {
// Set any requirement or hint on bundle according to its definition and
// uses. Return false if there are conflicting requirements which will
// require the bundle to be split.
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
VirtualRegister& reg = range->vreg();
if (range->hasDefinition()) {
// Deal with any definition constraints/hints.
LDefinition::Policy policy = reg.def()->policy();
if (policy == LDefinition::FIXED || policy == LDefinition::STACK) {
// Fixed and stack policies get a FIXED requirement. (In the stack
// case, the allocation should have been performed already by
// mergeAndQueueRegisters.)
JitSpewIfEnabled(JitSpew_RegAlloc,
" Requirement %s, fixed by definition",
reg.def()->output()->toString().get());
if (!requirement->merge(Requirement(*reg.def()->output()))) {
return false;
}
}
else if (reg.ins()->isPhi()) {
// Phis don't have any requirements, but they should prefer their
// input allocations. This is captured by the group hints above.
}
else {
// Non-phis get a REGISTER requirement.
if (!requirement->merge(Requirement(Requirement::
REGISTER))) {
return false;
}
}
}
// Search uses for requirements.
for (UsePositionIterator iter = range->usesBegin(); iter; iter++) {
LUse::Policy policy = iter->usePolicy();
if (policy == LUse::FIXED) {
AnyRegister required = GetFixedRegister(reg.def(), iter->use());
JitSpewIfEnabled(JitSpew_RegAlloc,
" Requirement %s, due to use at %u",
required.name(), iter->pos.bits());
// If there are multiple fixed registers which the bundle is
// required to use, fail. The bundle will need to be split before
// it can be allocated.
if (!requirement->merge(Requirement(LAllocation(required)))) {
return false;
}
}
else if (policy == LUse::
REGISTER) {
if (!requirement->merge(Requirement(Requirement::
REGISTER))) {
return false;
}
}
else if (policy == LUse::ANY) {
// ANY differs from KEEPALIVE by actively preferring a register.
if (!hint->merge(Requirement(Requirement::
REGISTER))) {
return false;
}
}
// The only case of STACK use policies is individual stack results using
// their containing stack result area, which is given a fixed allocation
// above.
MOZ_ASSERT_IF(policy == LUse::STACK,
requirement->kind() == Requirement::FIXED);
MOZ_ASSERT_IF(policy == LUse::STACK,
requirement->allocation().isStackArea());
}
}
return true;
}
bool BacktrackingAllocator::tryAllocateRegister(PhysicalRegister& r,
LiveBundle* bundle,
bool* success,
bool* hasCall,
LiveBundleVector& conflicting) {
*success =
false;
// If we know this bundle contains a call (where all registers are spilled) we
// shouldn't try again to allocate a register.
MOZ_ASSERT(!*hasCall);
if (!r.allocatable) {
return true;
}
LiveBundleVector aliasedConflicting;
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
LiveRangePlus rangePlus(range);
// All ranges in the bundle must be compatible with the physical register.
MOZ_ASSERT(range->vreg().isCompatible(r.reg));
const size_t numAliased = r.reg.numAliased();
for (size_t a = 0; a < numAliased; a++) {
PhysicalRegister& rAlias = registers[r.reg.aliased(a).code()];
LiveRangePlus existingPlus;
if (!rAlias.allocations.contains(rangePlus, &existingPlus)) {
continue;
}
const LiveRange* existing = existingPlus.liveRange();
if (existing->hasVreg()) {
MOZ_ASSERT(existing->bundle()->allocation().toRegister() == rAlias.reg);
bool duplicate =
false;
for (size_t i = 0; i < aliasedConflicting.length(); i++) {
if (aliasedConflicting[i] == existing->bundle()) {
duplicate =
true;
break;
}
}
if (!duplicate && !aliasedConflicting.append(existing->bundle())) {
return false;
}
}
else {
// This bundle contains a call instruction.
MOZ_ASSERT(lookupFirstCallPositionInRange(range->from(), range->to()));
JitSpewIfEnabled(JitSpew_RegAlloc,
" %s collides with fixed use %s",
rAlias.reg.name(), existing->toString().get());
*hasCall =
true;
return true;
}
MOZ_ASSERT(r.reg.numAliased() == numAliased);
}
}
if (!aliasedConflicting.empty()) {
// One or more aliased registers is allocated to another bundle
// overlapping this one. Keep track of the conflicting set, and in the
// case of multiple conflicting sets keep track of the set with the
// lowest maximum spill weight.
// The #ifdef guards against "unused variable 'existing'" bustage.
#ifdef JS_JITSPEW
if (JitSpewEnabled(JitSpew_RegAlloc)) {
if (aliasedConflicting.length() == 1) {
LiveBundle* existing = aliasedConflicting[0];
JitSpew(JitSpew_RegAlloc,
" %s collides with %s [weight %zu]",
r.reg.name(), existing->toString().get(),
computeSpillWeight(existing));
}
else {
JitSpew(JitSpew_RegAlloc,
" %s collides with the following",
r.reg.name());
for (size_t i = 0; i < aliasedConflicting.length(); i++) {
LiveBundle* existing = aliasedConflicting[i];
JitSpew(JitSpew_RegAlloc,
" %s [weight %zu]",
existing->toString().get(), computeSpillWeight(existing));
}
}
}
#endif
if (conflicting.empty()) {
conflicting = std::move(aliasedConflicting);
}
else {
if (maximumSpillWeight(aliasedConflicting) <
maximumSpillWeight(conflicting)) {
conflicting = std::move(aliasedConflicting);
}
}
return true;
}
JitSpewIfEnabled(JitSpew_RegAlloc,
" allocated to %s", r.reg.name());
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (!alloc().ensureBallast()) {
return false;
}
LiveRangePlus rangePlus(range);
if (!r.allocations.insert(rangePlus)) {
return false;
}
}
bundle->setAllocation(LAllocation(r.reg));
*success =
true;
return true;
}
bool BacktrackingAllocator::tryAllocateAnyRegister(
LiveBundle* bundle,
bool* success,
bool* hasCall,
LiveBundleVector& conflicting) {
// Search for any available register which the bundle can be allocated to.
LDefinition::Type type = bundle->firstRange()->vreg().type();
if (LDefinition::isFloatReg(type)) {
for (size_t i = AnyRegister::FirstFloatReg; i < AnyRegister::Total; i++) {
if (!LDefinition::isFloatRegCompatible(type, registers[i].reg.fpu())) {
continue;
}
if (!tryAllocateRegister(registers[i], bundle, success, hasCall,
conflicting)) {
return false;
}
if (*success) {
break;
}
if (*hasCall) {
// This bundle contains a call instruction. Calls require spilling all
// registers, so we have to split or spill this bundle.
break;
}
}
return true;
}
for (size_t i = 0; i < AnyRegister::FirstFloatReg; i++) {
if (!tryAllocateRegister(registers[i], bundle, success, hasCall,
conflicting)) {
return false;
}
if (*success) {
break;
}
if (*hasCall) {
// This bundle contains a call instruction. Calls require spilling all
// registers, so we have to split or spill this bundle.
break;
}
}
return true;
}
bool BacktrackingAllocator::evictBundle(LiveBundle* bundle) {
JitSpewIfEnabled(JitSpew_RegAlloc,
" Evicting %s [priority %zu] [weight %zu]",
bundle->toString().get(), computePriority(bundle),
computeSpillWeight(bundle));
AnyRegister reg(bundle->allocation().toRegister());
PhysicalRegister& physical = registers[reg.code()];
MOZ_ASSERT(physical.reg == reg && physical.allocatable);
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
LiveRangePlus rangePlus(range);
physical.allocations.remove(rangePlus);
}
bundle->setAllocation(LAllocation());
size_t priority = computePriority(bundle);
return allocationQueue.insert(QueueItem(bundle, priority));
}
bool BacktrackingAllocator::tryAllocateFixed(LiveBundle* bundle,
Requirement requirement,
bool* success,
bool* hasCall,
LiveBundleVector& conflicting) {
// Spill bundles which are required to be in a certain stack slot.
if (!requirement.allocation().isRegister()) {
JitSpew(JitSpew_RegAlloc,
" stack allocation requirement");
bundle->setAllocation(requirement.allocation());
*success =
true;
return true;
}
AnyRegister reg = requirement.allocation().toRegister();
return tryAllocateRegister(registers[reg.code()], bundle, success, hasCall,
conflicting);
}
bool BacktrackingAllocator::tryAllocateNonFixed(LiveBundle* bundle,
Requirement requirement,
Requirement hint,
bool* success,
bool* hasCall,
LiveBundleVector& conflicting) {
MOZ_ASSERT(hint.kind() != Requirement::FIXED);
MOZ_ASSERT(conflicting.empty());
// Spill bundles which have no hint or register requirement.
if (requirement.kind() == Requirement::NONE &&
hint.kind() == Requirement::NONE) {
JitSpew(JitSpew_RegAlloc,
" postponed spill (no hint or register requirement)");
if (!spilledBundles.append(bundle)) {
return false;
}
*success =
true;
return true;
}
if (!tryAllocateAnyRegister(bundle, success, hasCall, conflicting)) {
return false;
}
if (*success) {
return true;
}
// Spill bundles which have no register requirement if they didn't get
// allocated.
if (requirement.kind() == Requirement::NONE) {
JitSpew(JitSpew_RegAlloc,
" postponed spill (no register requirement)");
if (!spilledBundles.append(bundle)) {
return false;
}
*success =
true;
return true;
}
// We failed to allocate this bundle.
MOZ_ASSERT(!*success);
return true;
}
bool BacktrackingAllocator::processBundle(
const MIRGenerator* mir,
LiveBundle* bundle) {
JitSpewIfEnabled(JitSpew_RegAlloc,
"Allocating %s [priority %zu] [weight %zu]",
bundle->toString().get(), computePriority(bundle),
computeSpillWeight(bundle));
// A bundle can be processed by doing any of the following:
//
// - Assigning the bundle a register. The bundle cannot overlap any other
// bundle allocated for that physical register.
//
// - Spilling the bundle, provided it has no register uses.
//
// - Splitting the bundle into two or more bundles which cover the original
// one. The new bundles are placed back onto the priority queue for later
// processing.
//
// - Evicting one or more existing allocated bundles, and then doing one
// of the above operations. Evicted bundles are placed back on the
// priority queue. Any evicted bundles must have a lower spill weight
// than the bundle being processed.
//
// As long as this structure is followed, termination is guaranteed.
// In general, we want to minimize the amount of bundle splitting (which
// generally necessitates spills), so allocate longer lived, lower weight
// bundles first and evict and split them later if they prevent allocation
// for higher weight bundles.
Requirement requirement, hint;
bool doesNotHaveFixedConflict =
computeRequirement(bundle, &requirement, &hint);
bool hasCall =
false;
LiveBundleVector conflicting;
if (doesNotHaveFixedConflict) {
for (size_t attempt = 0;; attempt++) {
if (mir->shouldCancel(
"Backtracking Allocation (processBundle loop)")) {
return false;
}
bool success =
false;
hasCall =
false;
conflicting.clear();
// Ok, let's try allocating for this bundle.
if (requirement.kind() == Requirement::FIXED) {
if (!tryAllocateFixed(bundle, requirement, &success, &hasCall,
conflicting)) {
return false;
}
}
else {
if (!tryAllocateNonFixed(bundle, requirement, hint, &success, &hasCall,
conflicting)) {
return false;
}
}
// If that worked, we're done!
if (success) {
return true;
}
// If that didn't work, but we have one or more non-call bundles known to
// be conflicting, maybe we can evict them and try again.
if ((attempt < MAX_ATTEMPTS || minimalBundle(bundle)) && !hasCall &&
!conflicting.empty() &&
maximumSpillWeight(conflicting) < computeSpillWeight(bundle)) {
for (size_t i = 0; i < conflicting.length(); i++) {
if (!evictBundle(conflicting[i])) {
return false;
}
}
continue;
}
// We have to split this bundle.
break;
}
}
// A minimal bundle cannot be split any further. If we try to split it
// it at this point we will just end up with the same bundle and will
// enter an infinite loop. Weights and the initial live ranges must
// be constructed so that any minimal bundle is allocatable.
MOZ_ASSERT(!minimalBundle(bundle));
LiveBundle* conflict = conflicting.empty() ? nullptr : conflicting[0];
return chooseBundleSplit(bundle, hasCall, conflict);
}
// Helper for ::tryAllocatingRegistersForSpillBundles
bool BacktrackingAllocator::spill(LiveBundle* bundle) {
JitSpew(JitSpew_RegAlloc,
" Spilling bundle");
MOZ_ASSERT(bundle->allocation().isBogus());
if (LiveBundle* spillParent = bundle->spillParent()) {
JitSpew(JitSpew_RegAlloc,
" Using existing spill bundle");
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
LiveRange* parentRange = spillParent->rangeFor(range->from());
MOZ_ASSERT(parentRange->contains(range));
MOZ_ASSERT(&range->vreg() == &parentRange->vreg());
range->tryToMoveDefAndUsesInto(parentRange);
}
bundle->removeAllRangesFromVirtualRegisters();
return true;
}
return bundle->spillSet()->addSpilledBundle(bundle);
}
bool BacktrackingAllocator::tryAllocatingRegistersForSpillBundles() {
for (
auto it = spilledBundles.begin(); it != spilledBundles.end(); it++) {
LiveBundle* bundle = *it;
LiveBundleVector conflicting;
bool hasCall =
false;
bool success =
false;
if (mir->shouldCancel(
"Backtracking Try Allocating Spilled Bundles")) {
return false;
}
JitSpewIfEnabled(JitSpew_RegAlloc,
"Spill or allocate %s",
bundle->toString().get());
if (!tryAllocateAnyRegister(bundle, &success, &hasCall, conflicting)) {
return false;
}
// If the bundle still has no register, spill the bundle.
if (!success && !spill(bundle)) {
return false;
}
}
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Rewriting of the LIR after bundle processing is done: //
// ::pickStackSlots //
// ::createMoveGroupsFromLiveRangeTransitions //
// ::installAllocationsInLIR //
// ::populateSafepoints //
// ::annotateMoveGroups //
// //
///////////////////////////////////////////////////////////////////////////////
// Helper for ::pickStackSlot
bool BacktrackingAllocator::insertAllRanges(LiveRangePlusSet& set,
LiveBundle* bundle) {
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (!alloc().ensureBallast()) {
return false;
}
LiveRangePlus rangePlus(range);
if (!set.insert(rangePlus)) {
return false;
}
}
return true;
}
void BacktrackingAllocator::sortVirtualRegisterRanges() {
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
reg.sortRanges();
}
}
// Helper for ::pickStackSlots
bool BacktrackingAllocator::pickStackSlot(SpillSet* spillSet) {
// Look through all ranges that have been spilled in this set for a
// register definition which is fixed to a stack or argument slot. If we
// find one, use it for all bundles that have been spilled. tryMergeBundles
// makes sure this reuse is possible when an initial bundle contains ranges
// from multiple virtual registers.
for (size_t i = 0; i < spillSet->numSpilledBundles(); i++) {
LiveBundle* bundle = spillSet->spilledBundle(i);
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter; iter++) {
LiveRange* range = *iter;
if (range->hasDefinition()) {
LDefinition* def = range->vreg().def();
if (def->policy() == LDefinition::FIXED) {
MOZ_ASSERT(!def->output()->isRegister());
MOZ_ASSERT(!def->output()->isStackSlot());
spillSet->setAllocation(*def->output());
return true;
}
}
}
}
LDefinition::Type type =
spillSet->spilledBundle(0)->firstRange()->vreg().type();
SpillSlotList* slotList;
switch (StackSlotAllocator::width(type)) {
case 4:
slotList = &normalSlots;
break;
case 8:
slotList = &doubleSlots;
break;
case 16:
slotList = &quadSlots;
break;
default:
MOZ_CRASH(
"Bad width");
}
// Maximum number of existing spill slots we will look at before giving up
// and allocating a new slot.
static const size_t MAX_SEARCH_COUNT = 10;
size_t searches = 0;
SpillSlot* stop = nullptr;
while (!slotList->empty()) {
SpillSlot* spillSlot = *slotList->begin();
if (!stop) {
stop = spillSlot;
}
else if (stop == spillSlot) {
// We looked through every slot in the list.
break;
}
bool success =
true;
for (size_t i = 0; i < spillSet->numSpilledBundles(); i++) {
LiveBundle* bundle = spillSet->spilledBundle(i);
for (LiveBundle::RangeIterator iter = bundle->rangesBegin(); iter;
iter++) {
LiveRange* range = *iter;
LiveRangePlus rangePlus(range);
LiveRangePlus existingPlus;
if (spillSlot->allocated.contains(rangePlus, &existingPlus)) {
success =
false;
break;
}
}
if (!success) {
break;
}
}
if (success) {
// We can reuse this physical stack slot for the new bundles.
// Update the allocated ranges for the slot.
for (size_t i = 0; i < spillSet->numSpilledBundles(); i++) {
LiveBundle* bundle = spillSet->spilledBundle(i);
if (!insertAllRanges(spillSlot->allocated, bundle)) {
return false;
}
}
spillSet->setAllocation(spillSlot->alloc);
return true;
}
// On a miss, move the spill to the end of the list. This will cause us
// to make fewer attempts to allocate from slots with a large and
// highly contended range.
slotList->popFront();
slotList->pushBack(spillSlot);
if (++searches == MAX_SEARCH_COUNT) {
break;
}
}
// We need a new physical stack slot.
uint32_t stackSlot = stackSlotAllocator.allocateSlot(type);
SpillSlot* spillSlot =
new (alloc().fallible()) SpillSlot(stackSlot, alloc().lifoAlloc());
if (!spillSlot) {
return false;
}
for (size_t i = 0; i < spillSet->numSpilledBundles(); i++) {
LiveBundle* bundle = spillSet->spilledBundle(i);
if (!insertAllRanges(spillSlot->allocated, bundle)) {
return false;
}
}
spillSet->setAllocation(spillSlot->alloc);
slotList->pushFront(spillSlot);
return true;
}
bool BacktrackingAllocator::pickStackSlots() {
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (mir->shouldCancel(
"Backtracking Pick Stack Slots")) {
return false;
}
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
LiveBundle* bundle = iter->bundle();
if (bundle->allocation().isBogus()) {
if (!pickStackSlot(bundle->spillSet())) {
return false;
}
MOZ_ASSERT(!bundle->allocation().isBogus());
}
}
}
return true;
}
// Helper for ::createMoveGroupsFromLiveRangeTransitions
bool BacktrackingAllocator::moveAtEdge(LBlock* predecessor, LBlock* successor,
LiveRange* from, LiveRange* to,
LDefinition::Type type) {
if (successor->mir()->numPredecessors() > 1) {
MOZ_ASSERT(predecessor->mir()->numSuccessors() == 1);
return moveAtExit(predecessor, from, to, type);
}
return moveAtEntry(successor, from, to, type);
}
// Helper for ::createMoveGroupsFromLiveRangeTransitions
void BacktrackingAllocator::removeDeadRanges(VirtualRegister& reg) {
auto isDeadRange = [&](VirtualRegister::RangeVector& ranges,
LiveRange* range) {
// Check for direct uses of this range.
if (range->hasUses() || range->hasDefinition()) {
return false;
}
// Check if there are later ranges for this vreg. The first item in the list
// must have the highest start position so we compare against that one.
CodePosition start = range->from();
CodePosition maxFrom = ranges[0]->from();
if (maxFrom > start) {
return false;
}
LNode* ins = insData[start];
if (start == entryOf(ins->block())) {
return false;
}
// Check if this range ends at a loop backedge.
LNode* last = insData[range->to().previous()];
if (last->isGoto() &&
last->toGoto()->target()->id() < last->block()->mir()->id()) {
return false;
}
// Check if there are phis which this vreg flows to.
if (reg.usedByPhi()) {
return false;
}
return true;
};
reg.removeRangesIf(isDeadRange);
}
static void AssertCorrectRangeForPosition(
const VirtualRegister& reg,
CodePosition pos,
const LiveRange* range) {
MOZ_ASSERT(range->covers(pos));
#ifdef DEBUG
// Assert the result is consistent with rangeFor. The ranges can be different
// but must be equivalent (both register or both non-register ranges).
LiveRange* expected = reg.rangeFor(pos,
/* preferRegister = */ true);
MOZ_ASSERT(range->bundle()->allocation().isRegister() ==
expected->bundle()->allocation().isRegister());
#endif
}
// Helper for ::createMoveGroupsFromLiveRangeTransitions
bool BacktrackingAllocator::createMoveGroupsForControlFlowEdges(
const VirtualRegister& reg,
const ControlFlowEdgeVector& edges) {
// Iterate over both the virtual register ranges (sorted by start position)
// and the control flow edges (sorted by predecessorExit). When we find the
// predecessor range for the next edge, add a move from predecessor range to
// successor range.
VirtualRegister::RangeIterator iter(reg);
LiveRange* nonRegisterRange = nullptr;
for (
const ControlFlowEdge& edge : edges) {
CodePosition pos = edge.predecessorExit;
LAllocation successorAllocation =
edge.successorRange->bundle()->allocation();
// We don't need to insert a move if nonRegisterRange covers the predecessor
// block exit and has the same allocation as the successor block. This check
// is not required for correctness but it reduces the number of generated
// moves.
if (nonRegisterRange && pos < nonRegisterRange->to() &&
nonRegisterRange->bundle()->allocation() == successorAllocation) {
MOZ_ASSERT(nonRegisterRange->covers(pos));
continue;
}
// Search for a matching range. Prefer a register range.
LiveRange* predecessorRange = nullptr;
bool foundSameAllocation =
false;
while (
true) {
if (iter.done() || iter->from() > pos) {
// No register range covers this edge.
predecessorRange = nonRegisterRange;
break;
}
if (iter->to() <= pos) {
// Skip ranges that end before this edge (and later edges).
iter++;
continue;
}
MOZ_ASSERT(iter->covers(pos));
if (iter->bundle()->allocation() == successorAllocation) {
// There's a range covering the predecessor block exit that has the same
// allocation, so we don't need to insert a move. This check is not
// required for correctness but it reduces the number of generated
// moves.
foundSameAllocation =
true;
break;
}
if (iter->bundle()->allocation().isRegister()) {
predecessorRange = *iter;
break;
}
if (!nonRegisterRange || iter->to() > nonRegisterRange->to()) {
nonRegisterRange = *iter;
}
iter++;
}
if (foundSameAllocation) {
continue;
}
MOZ_ASSERT(predecessorRange);
AssertCorrectRangeForPosition(reg, pos, predecessorRange);
if (!alloc().ensureBallast()) {
return false;
}
JitSpew(JitSpew_RegAlloc,
" (moveAtEdge#2)");
if (!moveAtEdge(edge.predecessor, edge.successor, predecessorRange,
edge.successorRange, reg.type())) {
return false;
}
}
return true;
}
bool BacktrackingAllocator::createMoveGroupsFromLiveRangeTransitions() {
// Add moves to handle changing assignments for vregs over their lifetime.
JitSpew(JitSpew_RegAlloc,
"ResolveControlFlow: begin");
JitSpew(JitSpew_RegAlloc,
" ResolveControlFlow: adding MoveGroups within blocks");
// Look for places where a register's assignment changes in the middle of a
// basic block.
MOZ_ASSERT(!vregs[0u].hasRanges());
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (mir->shouldCancel(
"Backtracking Resolve Control Flow (vreg outer loop)")) {
return false;
}
// Remove ranges which will never be used.
removeDeadRanges(reg);
LiveRange* registerRange = nullptr;
LiveRange* nonRegisterRange = nullptr;
VirtualRegister::RangeIterator iter(reg);
// Keep track of the register and non-register ranges with the highest end
// position before advancing the iterator. These are predecessor ranges for
// later ranges.
auto moveToNextRange = [&](LiveRange* range) {
MOZ_ASSERT(*iter == range);
if (range->bundle()->allocation().isRegister()) {
if (!registerRange || range->to() > registerRange->to()) {
registerRange = range;
}
}
else {
if (!nonRegisterRange || range->to() > nonRegisterRange->to()) {
nonRegisterRange = range;
}
}
iter++;
};
// Iterate over all ranges.
while (!iter.done()) {
LiveRange* range = *iter;
if (mir->shouldCancel(
"Backtracking Resolve Control Flow (vreg inner loop)")) {
return false;
}
// The range which defines the register does not have a predecessor
// to add moves from.
if (range->hasDefinition()) {
moveToNextRange(range);
continue;
}
// Ignore ranges that start at block boundaries. We will handle
// these in the next phase.
CodePosition start = range->from();
LNode* ins = insData[start];
if (start == entryOf(ins->block())) {
moveToNextRange(range);
continue;
}
// Determine the predecessor range to use for this range and other ranges
// starting at the same position. Prefer a register range.
LiveRange* predecessorRange = nullptr;
if (registerRange && start.previous() < registerRange->to()) {
predecessorRange = registerRange;
}
else {
MOZ_ASSERT(nonRegisterRange);
MOZ_ASSERT(start.previous() < nonRegisterRange->to());
predecessorRange = nonRegisterRange;
}
AssertCorrectRangeForPosition(reg, start.previous(), predecessorRange);
// Add moves from predecessorRange to all ranges that start here.
do {
range = *iter;
MOZ_ASSERT(!range->hasDefinition());
if (!alloc().ensureBallast()) {
return false;
}
#ifdef DEBUG
// If we already saw a range which covers the start of this range, it
// must have a different allocation.
for (VirtualRegister::RangeIterator prevIter(reg); *prevIter != range;
prevIter++) {
MOZ_ASSERT_IF(prevIter->covers(start),
prevIter->bundle()->allocation() !=
range->bundle()->allocation());
}
#endif
if (start.subpos() == CodePosition::INPUT) {
JitSpewIfEnabled(JitSpew_RegAlloc,
" moveInput (%s) <- (%s)",
range->toString().get(),
predecessorRange->toString().get());
if (!moveInput(ins->toInstruction(), predecessorRange, range,
reg.type())) {
return false;
}
}
else {
JitSpew(JitSpew_RegAlloc,
" (moveAfter)");
if (!moveAfter(ins->toInstruction(), predecessorRange, range,
reg.type())) {
return false;
}
}
moveToNextRange(range);
}
while (!iter.done() && iter->from() == start);
}
}
JitSpew(JitSpew_RegAlloc,
" ResolveControlFlow: adding MoveGroups for phi nodes");
for (size_t i = 0; i < graph.numBlocks(); i++) {
if (mir->shouldCancel(
"Backtracking Resolve Control Flow (block loop)")) {
return false;
}
LBlock* successor = graph.getBlock(i);
MBasicBlock* mSuccessor = successor->mir();
if (mSuccessor->numPredecessors() < 1) {
continue;
}
// Resolve phis to moves.
for (size_t j = 0; j < successor->numPhis(); j++) {
LPhi* phi = successor->getPhi(j);
MOZ_ASSERT(phi->numDefs() == 1);
LDefinition* def = phi->getDef(0);
VirtualRegister& reg = vreg(def);
LiveRange* to = reg.firstRange();
MOZ_ASSERT(to->from() == entryOf(successor));
for (size_t k = 0; k < mSuccessor->numPredecessors(); k++) {
LBlock* predecessor = mSuccessor->getPredecessor(k)->lir();
MOZ_ASSERT(predecessor->mir()->numSuccessors() == 1);
LAllocation* input = phi->getOperand(k);
LiveRange* from = vreg(input).rangeFor(exitOf(predecessor),
/* preferRegister = */ true);
MOZ_ASSERT(from);
if (!alloc().ensureBallast()) {
return false;
}
// Note: we have to use moveAtEdge both here and below (for edge
// resolution) to avoid conflicting moves. See bug 1493900.
JitSpew(JitSpew_RegAlloc,
" (moveAtEdge#1)");
if (!moveAtEdge(predecessor, successor, from, to, def->type())) {
return false;
}
}
}
}
JitSpew(JitSpew_RegAlloc,
" ResolveControlFlow: adding MoveGroups to fix conflicted edges");
// Add moves to resolve graph edges with different allocations at their
// source and target.
ControlFlowEdgeVector edges;
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
// First collect all control flow edges we need to resolve. This loop knows
// the range on the successor side, but looking up the corresponding
// predecessor range with rangeFor is quadratic so we handle that
// differently.
edges.clear();
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
LiveRange* targetRange = *iter;
size_t firstBlockId = insData[targetRange->from()]->block()->mir()->id();
if (!targetRange->covers(entryOf(graph.getBlock(firstBlockId)))) {
firstBlockId++;
}
for (size_t id = firstBlockId; id < graph.numBlocks(); id++) {
LBlock* successor = graph.getBlock(id);
if (!targetRange->covers(entryOf(successor))) {
break;
}
VirtualRegBitSet& live = liveIn[id];
if (!live.contains(i)) {
continue;
}
for (size_t j = 0; j < successor->mir()->numPredecessors(); j++) {
LBlock* predecessor = successor->mir()->getPredecessor(j)->lir();
CodePosition predecessorExit = exitOf(predecessor);
if (targetRange->covers(predecessorExit)) {
continue;
}
if (!edges.emplaceBack(predecessor, successor, targetRange,
predecessorExit)) {
return false;
}
}
}
}
if (edges.empty()) {
continue;
}
// Sort edges by predecessor position. This doesn't need to be a stable sort
// because createMoveGroupsForControlFlowEdges will use the same predecessor
// range if there are multiple edges with the same predecessor position.
auto compareEdges = [](
const ControlFlowEdge& a,
const ControlFlowEdge& b) {
return a.predecessorExit < b.predecessorExit;
};
std::sort(edges.begin(), edges.end(), compareEdges);
// Resolve edges and add move groups.
if (!createMoveGroupsForControlFlowEdges(reg, edges)) {
return false;
}
}
JitSpew(JitSpew_RegAlloc,
"ResolveControlFlow: end");
return true;
}
// Helper for ::addLiveRegistersForRange
size_t BacktrackingAllocator::findFirstNonCallSafepoint(CodePosition pos,
size_t startFrom) {
// Assert startFrom is valid.
MOZ_ASSERT_IF(startFrom > 0,
inputOf(graph.getSafepoint(startFrom - 1)) < pos);
size_t i = startFrom;
for (; i < graph.numNonCallSafepoints(); i++) {
const LInstruction* ins = graph.getNonCallSafepoint(i);
if (pos <= inputOf(ins)) {
break;
}
}
return i;
}
// Helper for ::installAllocationsInLIR
void BacktrackingAllocator::addLiveRegistersForRange(
VirtualRegister& reg, LiveRange* range, size_t* firstNonCallSafepoint) {
// Fill in the live register sets for all non-call safepoints.
LAllocation a = range->bundle()->allocation();
if (!a.isRegister()) {
return;
}
// Don't add output registers to the safepoint.
CodePosition start = range->from();
if (range->hasDefinition() && !reg.isTemp()) {
#ifdef CHECK_OSIPOINT_REGISTERS
// We don't add the output register to the safepoint,
// but it still might get added as one of the inputs.
// So eagerly add this reg to the safepoint clobbered registers.
if (reg.ins()->isInstruction()) {
if (LSafepoint* safepoint = reg.ins()->toInstruction()->safepoint()) {
safepoint->addClobberedRegister(a.toRegister());
}
}
#endif
start = start.next();
}
*firstNonCallSafepoint =
findFirstNonCallSafepoint(range->from(), *firstNonCallSafepoint);
for (size_t i = *firstNonCallSafepoint; i < graph.numNonCallSafepoints();
i++) {
LInstruction* ins = graph.getNonCallSafepoint(i);
CodePosition pos = inputOf(ins);
// Safepoints are sorted, so we can shortcut out of this loop
// if we go out of range.
if (range->to() <= pos) {
break;
}
MOZ_ASSERT(range->covers(pos));
LSafepoint* safepoint = ins->safepoint();
safepoint->addLiveRegister(a.toRegister());
#ifdef CHECK_OSIPOINT_REGISTERS
if (reg.isTemp()) {
safepoint->addClobberedRegister(a.toRegister());
}
#endif
}
}
// Helper for ::installAllocationsInLIR
static inline size_t NumReusingDefs(LInstruction* ins) {
size_t num = 0;
for (size_t i = 0; i < ins->numDefs(); i++) {
LDefinition* def = ins->getDef(i);
if (def->policy() == LDefinition::MUST_REUSE_INPUT) {
num++;
}
}
return num;
}
bool BacktrackingAllocator::installAllocationsInLIR() {
JitSpew(JitSpew_RegAlloc,
"Installing Allocations");
// The virtual registers and safepoints are both ordered by position. To avoid
// quadratic behavior in findFirstNonCallSafepoint, we use
// firstNonCallSafepoint as cursor to start the search at the safepoint
// returned by the previous call.
size_t firstNonCallSafepoint = 0;
MOZ_ASSERT(!vregs[0u].hasRanges());
for (size_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (mir->shouldCancel(
"Backtracking Install Allocations (main loop)")) {
return false;
}
firstNonCallSafepoint =
findFirstNonCallSafepoint(inputOf(reg.ins()), firstNonCallSafepoint);
// The ranges are sorted by start position, so we can use the same
// findFirstNonCallSafepoint optimization here.
size_t firstNonCallSafepointForRange = firstNonCallSafepoint;
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
LiveRange* range = *iter;
if (range->hasDefinition()) {
reg.def()->setOutput(range->bundle()->allocation());
if (reg.ins()->recoversInput()) {
LSnapshot* snapshot = reg.ins()->toInstruction()->snapshot();
for (size_t i = 0; i < snapshot->numEntries(); i++) {
LAllocation* entry = snapshot->getEntry(i);
if (entry->isUse() &&
entry->toUse()->policy() == LUse::RECOVERED_INPUT) {
*entry = *reg.def()->output();
}
}
}
}
for (UsePositionIterator iter(range->usesBegin()); iter; iter++) {
LAllocation* alloc = iter->use();
*alloc = range->bundle()->allocation();
// For any uses which feed into MUST_REUSE_INPUT definitions,
// add copies if the use and def have different allocations.
LNode* ins = insData[iter->pos];
if (LDefinition* def = FindReusingDefOrTemp(ins, alloc)) {
LiveRange* outputRange = vreg(def).firstRange();
MOZ_ASSERT(outputRange->covers(outputOf(ins)));
LAllocation res = outputRange->bundle()->allocation();
LAllocation sourceAlloc = range->bundle()->allocation();
if (res != *alloc) {
if (!this->alloc().ensureBallast()) {
return false;
}
if (NumReusingDefs(ins->toInstruction()) <= 1) {
LMoveGroup* group = getInputMoveGroup(ins->toInstruction());
if (!group->addAfter(sourceAlloc, res, reg.type())) {
return false;
}
}
else {
LMoveGroup* group = getFixReuseMoveGroup(ins->toInstruction());
if (!group->add(sourceAlloc, res, reg.type())) {
return false;
}
}
*alloc = res;
}
}
}
addLiveRegistersForRange(reg, range, &firstNonCallSafepointForRange);
}
}
graph.setLocalSlotsSize(stackSlotAllocator.stackHeight());
return true;
}
// Helper for ::populateSafepoints
size_t BacktrackingAllocator::findFirstSafepoint(CodePosition pos,
size_t startFrom) {
// Assert startFrom is valid.
MOZ_ASSERT_IF(startFrom > 0,
inputOf(graph.getSafepoint(startFrom - 1)) < pos);
size_t i = startFrom;
for (; i < graph.numSafepoints(); i++) {
LInstruction* ins = graph.getSafepoint(i);
if (pos <= inputOf(ins)) {
break;
}
}
return i;
}
// Helper for ::populateSafepoints
static inline bool IsNunbox(VirtualRegister& reg) {
#ifdef JS_NUNBOX32
return reg.type() == LDefinition::TYPE || reg.type() == LDefinition::PAYLOAD;
#else
return false;
#endif
}
// Helper for ::populateSafepoints
static inline bool IsSlotsOrElements(VirtualRegister& reg) {
return reg.type() == LDefinition::SLOTS;
}
// Helper for ::populateSafepoints
static inline bool IsTraceable(VirtualRegister& reg) {
if (reg.type() == LDefinition::OBJECT ||
reg.type() == LDefinition::WASM_ANYREF) {
return true;
}
#ifdef JS_PUNBOX64
if (reg.type() == LDefinition::BOX) {
return true;
}
#endif
if (reg.type() == LDefinition::STACKRESULTS) {
MOZ_ASSERT(reg.def());
const LStackArea* alloc = reg.def()->output()->toStackArea();
for (
auto iter = alloc->results(); iter; iter.next()) {
if (iter.isWasmAnyRef()) {
return true;
}
}
}
return false;
}
bool BacktrackingAllocator::populateSafepoints() {
JitSpew(JitSpew_RegAlloc,
"Populating Safepoints");
// The virtual registers and safepoints are both ordered by position. To avoid
// quadratic behavior in findFirstSafepoint, we use firstSafepoint as cursor
// to start the search at the safepoint returned by the previous call.
size_t firstSafepoint = 0;
MOZ_ASSERT(!vregs[0u].def());
for (uint32_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
if (!reg.def() ||
(!IsTraceable(reg) && !IsSlotsOrElements(reg) && !IsNunbox(reg))) {
continue;
}
firstSafepoint = findFirstSafepoint(inputOf(reg.ins()), firstSafepoint);
if (firstSafepoint >= graph.numSafepoints()) {
break;
}
// The ranges are sorted by start position, so we can use the same
// findFirstSafepoint optimization here.
size_t firstSafepointForRange = firstSafepoint;
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
LiveRange* range = *iter;
firstSafepointForRange =
findFirstSafepoint(range->from(), firstSafepointForRange);
for (size_t j = firstSafepointForRange; j < graph.numSafepoints(); j++) {
LInstruction* ins = graph.getSafepoint(j);
if (inputOf(ins) >= range->to()) {
break;
}
MOZ_ASSERT(range->covers(inputOf(ins)));
// Include temps but not instruction outputs. Also make sure
// MUST_REUSE_INPUT is not used with gcthings or nunboxes, or
// we would have to add the input reg to this safepoint.
if (ins == reg.ins() && !reg.isTemp()) {
DebugOnly<LDefinition*> def = reg.def();
MOZ_ASSERT_IF(def->policy() == LDefinition::MUST_REUSE_INPUT,
def->type() == LDefinition::GENERAL ||
def->type() == LDefinition::INT32 ||
def->type() == LDefinition::FLOAT32 ||
def->type() == LDefinition::
DOUBLE ||
def->type() == LDefinition::SIMD128);
continue;
}
LSafepoint* safepoint = ins->safepoint();
LAllocation a = range->bundle()->allocation();
if (a.isGeneralReg() && ins->isCall()) {
continue;
}
switch (reg.type()) {
case LDefinition::OBJECT:
if (!safepoint->addGcPointer(a)) {
return false;
}
break;
case LDefinition::SLOTS:
if (!safepoint->addSlotsOrElementsPointer(a)) {
return false;
}
break;
case LDefinition::WASM_ANYREF:
if (!safepoint->addWasmAnyRef(a)) {
return false;
}
break;
case LDefinition::STACKRESULTS: {
MOZ_ASSERT(a.isStackArea());
for (
auto iter = a.toStackArea()->results(); iter; iter.next()) {
if (iter.isWasmAnyRef()) {
if (!safepoint->addWasmAnyRef(iter.alloc())) {
return false;
}
}
}
break;
}
#ifdef JS_NUNBOX32
case LDefinition::TYPE:
if (!safepoint->addNunboxType(i, a)) {
return false;
}
break;
case LDefinition::PAYLOAD:
if (!safepoint->addNunboxPayload(i, a)) {
return false;
}
break;
#else
case LDefinition::BOX:
if (!safepoint->addBoxedValue(a)) {
return false;
}
break;
#endif
default:
MOZ_CRASH(
"Bad register type");
}
}
}
}
return true;
}
bool BacktrackingAllocator::annotateMoveGroups() {
// Annotate move groups in the LIR graph with any register that is not
// allocated at that point and can be used as a scratch register. This is
// only required for x86, as other platforms always have scratch registers
// available for use.
#ifdef JS_CODEGEN_X86
LiveRange* range = LiveRange::FallibleNew(alloc(), nullptr, CodePosition(),
CodePosition().next());
if (!range) {
return false;
}
for (size_t i = 0; i < graph.numBlocks(); i++) {
if (mir->shouldCancel(
"Backtracking Annotate Move Groups")) {
return false;
}
LBlock* block = graph.getBlock(i);
LInstruction* last = nullptr;
for (LInstructionIterator iter = block->begin(); iter != block->end();
++iter) {
if (iter->isMoveGroup()) {
CodePosition from = last ? outputOf(last) : entryOf(block);
range->setTo(from.next());
range->setFrom(from);
for (size_t i = 0; i < AnyRegister::Total; i++) {
PhysicalRegister& reg = registers[i];
if (reg.reg.isFloat() || !reg.allocatable) {
continue;
}
// This register is unavailable for use if (a) it is in use
// by some live range immediately before the move group,
// or (b) it is an operand in one of the group's moves. The
// latter case handles live ranges which end immediately
// before the move group or start immediately after.
// For (b) we need to consider move groups immediately
// preceding or following this one.
if (iter->toMoveGroup()->uses(reg.reg.gpr())) {
continue;
}
bool found =
false;
LInstructionIterator niter(iter);
for (niter++; niter != block->end(); niter++) {
if (niter->isMoveGroup()) {
if (niter->toMoveGroup()->uses(reg.reg.gpr())) {
found =
true;
break;
}
}
else {
break;
}
}
if (iter != block->begin()) {
LInstructionIterator riter(iter);
do {
riter--;
if (riter->isMoveGroup()) {
if (riter->toMoveGroup()->uses(reg.reg.gpr())) {
found =
true;
break;
}
}
else {
break;
}
}
while (riter != block->begin());
}
if (found) {
continue;
}
LiveRangePlus existingPlus;
LiveRangePlus rangePlus(range);
if (reg.allocations.contains(rangePlus, &existingPlus)) {
continue;
}
iter->toMoveGroup()->setScratchRegister(reg.reg.gpr());
break;
}
}
else {
last = *iter;
}
}
}
#endif
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
// Debug-printing support //
// //
///////////////////////////////////////////////////////////////////////////////
#ifdef JS_JITSPEW
UniqueChars LiveRange::toString()
const {
AutoEnterOOMUnsafeRegion oomUnsafe;
UniqueChars buf = JS_smprintf(
"v%u %u-%u", hasVreg() ? vreg().vreg() : 0,
from().bits(), to().bits() - 1);
if (buf && bundle() && !bundle()->allocation().isBogus()) {
buf = JS_sprintf_append(std::move(buf),
" %s",
bundle()->allocation().toString().get());
}
buf = JS_sprintf_append(std::move(buf),
" {");
if (buf && hasDefinition()) {
buf = JS_sprintf_append(std::move(buf),
" %u_def", from().bits());
if (hasVreg()) {
// If the definition has a fixed requirement, print it too.
const LDefinition* def = vreg().def();
LDefinition::Policy policy = def->policy();
if (policy == LDefinition::FIXED || policy == LDefinition::STACK) {
if (buf) {
buf = JS_sprintf_append(std::move(buf),
":F:%s",
def->output()->toString().get());
}
}
}
}
for (UsePositionIterator iter = usesBegin(); buf && iter; iter++) {
buf = JS_sprintf_append(std::move(buf),
" %u_%s", iter->pos.bits(),
iter->use()->toString().get());
}
buf = JS_sprintf_append(std::move(buf),
" }");
if (!buf) {
oomUnsafe.crash(
"LiveRange::toString()");
}
return buf;
}
UniqueChars LiveBundle::toString()
const {
AutoEnterOOMUnsafeRegion oomUnsafe;
UniqueChars buf = JS_smprintf(
"LB%u(", id());
if (buf) {
if (spillParent()) {
buf =
JS_sprintf_append(std::move(buf),
"parent=LB%u", spillParent()->id());
}
else {
buf = JS_sprintf_append(std::move(buf),
"parent=none");
}
}
for (LiveBundle::RangeIterator iter = rangesBegin(); buf && iter; iter++) {
if (buf) {
buf = JS_sprintf_append(std::move(buf),
"%s %s",
(iter == rangesBegin()) ?
"" :
" ##",
iter->toString().get());
}
}
if (buf) {
buf = JS_sprintf_append(std::move(buf),
")");
}
if (!buf) {
oomUnsafe.crash(
"LiveBundle::toString()");
}
return buf;
}
void BacktrackingAllocator::dumpLiveRangesByVReg(
const char* who) {
MOZ_ASSERT(!vregs[0u].hasRanges());
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Live ranges by virtual register (%s):", who);
for (uint32_t i = 1; i < graph.numVirtualRegisters(); i++) {
JitSpewHeader(JitSpew_RegAlloc);
JitSpewCont(JitSpew_RegAlloc,
" ");
VirtualRegister& reg = vregs[i];
for (VirtualRegister::RangeIterator iter(reg); iter; iter++) {
if (*iter != reg.firstRange()) {
JitSpewCont(JitSpew_RegAlloc,
" ## ");
}
JitSpewCont(JitSpew_RegAlloc,
"%s", iter->toString().get());
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
}
}
void BacktrackingAllocator::dumpLiveRangesByBundle(
const char* who) {
MOZ_ASSERT(!vregs[0u].hasRanges());
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Live ranges by bundle (%s):", who);
for (uint32_t i = 1; i < graph.numVirtualRegisters(); i++) {
VirtualRegister& reg = vregs[i];
for (VirtualRegister::RangeIterator baseIter(reg); baseIter; baseIter++) {
LiveRange* range = *baseIter;
LiveBundle* bundle = range->bundle();
if (range == bundle->firstRange()) {
JitSpew(JitSpew_RegAlloc,
" %s", bundle->toString().get());
}
}
}
}
void BacktrackingAllocator::dumpAllocations() {
JitSpew(JitSpew_RegAlloc,
"Allocations:");
dumpLiveRangesByBundle(
"in dumpAllocations()");
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Allocations by physical register:");
for (size_t i = 0; i < AnyRegister::Total; i++) {
if (registers[i].allocatable && !registers[i].allocations.empty()) {
JitSpewHeader(JitSpew_RegAlloc);
JitSpewCont(JitSpew_RegAlloc,
" %s:", AnyRegister::FromCode(i).name());
bool first =
true;
LiveRangePlusSet::Iter lrpIter(®isters[i].allocations);
while (lrpIter.hasMore()) {
LiveRange* range = lrpIter.next().liveRange();
if (first) {
first =
false;
}
else {
fprintf(stderr,
" /");
}
fprintf(stderr,
" %s", range->toString().get());
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
}
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
}
#endif // JS_JITSPEW
///////////////////////////////////////////////////////////////////////////////
// //
// Top level of the register allocation machinery //
// //
///////////////////////////////////////////////////////////////////////////////
bool BacktrackingAllocator::go() {
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Beginning register allocation");
JitSpewCont(JitSpew_RegAlloc,
"\n");
if (JitSpewEnabled(JitSpew_RegAlloc)) {
dumpInstructions(
"(Pre-allocation LIR)");
}
if (!init()) {
return false;
}
if (!buildLivenessInfo()) {
return false;
}
#ifdef JS_JITSPEW
if (JitSpewEnabled(JitSpew_RegAlloc)) {
dumpLiveRangesByVReg(
"after liveness analysis");
}
#endif
if (!allocationQueue.reserve(graph.numVirtualRegisters() * 3 / 2)) {
return false;
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Beginning grouping and queueing registers");
if (!mergeAndQueueRegisters()) {
return false;
}
JitSpew(JitSpew_RegAlloc,
"Completed grouping and queueing registers");
#ifdef JS_JITSPEW
if (JitSpewEnabled(JitSpew_RegAlloc)) {
dumpLiveRangesByBundle(
"after grouping/queueing regs");
}
#endif
// There now follow two allocation loops, which are really the heart of the
// allocator. First, the "main" allocation loop. This does almost all of
// the allocation work, by repeatedly pulling bundles out of
// ::allocationQueue and calling ::processBundle on it, until there are no
// bundles left in the queue. Note that ::processBundle can add new smaller
// bundles to the queue if it needs to split or spill a bundle.
//
// For each bundle in turn pulled out of ::allocationQueue, ::processBundle:
//
// * calls ::computeRequirement to discover the overall constraint for the
// bundle.
//
// * tries to find a register for it, by calling either ::tryAllocateFixed or
// ::tryAllocateNonFixed.
//
// * if that fails, but ::tryAllocateFixed / ::tryAllocateNonFixed indicate
// that there is some other bundle with lower spill weight that can be
// evicted, then that bundle is evicted (hence, put back into
// ::allocationQueue), and we try again.
//
// * at most MAX_ATTEMPTS may be made.
//
// * If that still fails to find a register, then the bundle is handed off to
// ::chooseBundleSplit. That will choose to either split the bundle,
// yielding multiple pieces which are put back into ::allocationQueue, or
// it will spill the bundle. Note that the same mechanism applies to both;
// there's no clear boundary between splitting and spilling, because
// spilling can be interpreted as an extreme form of splitting.
//
// ::processBundle and its callees contains much gnarly and logic which isn't
// easy to understand, particularly in the area of how eviction candidates
// are chosen. But it works well enough, and tinkering doesn't seem to
// improve the resulting allocations. More important is the splitting logic,
// because that controls where spill/reload instructions are placed.
//
// Eventually ::allocationQueue becomes empty, and each LiveBundle has either
// been allocated a register or is marked for spilling. In the latter case
// it will have been added to ::spilledBundles.
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Beginning main allocation loop");
JitSpewCont(JitSpew_RegAlloc,
"\n");
// Allocate, spill and split bundles until finished.
while (!allocationQueue.empty()) {
if (mir->shouldCancel(
"Backtracking Allocation")) {
return false;
}
QueueItem item = allocationQueue.removeHighest();
if (!processBundle(mir, item.bundle)) {
return false;
}
}
// And here's the second allocation loop (hidden inside
// ::tryAllocatingRegistersForSpillBundles). It makes one last attempt to
// find a register for each spill bundle. There's no attempt to free up
// registers by eviction. In at least 99% of cases this attempt fails, in
// which case the bundle is handed off to ::spill. The lucky remaining 1%
// get a register. Unfortunately this scheme interacts badly with the
// splitting strategy, leading to excessive register-to-register copying in
// some very simple cases. See bug 1752520.
//
// A modest but probably worthwhile amount of allocation time can be saved by
// making ::tryAllocatingRegistersForSpillBundles use specialised versions of
// ::tryAllocateAnyRegister and its callees, that don't bother to create sets
// of conflicting bundles. Creating those sets is expensive and, here,
// pointless, since we're not going to do any eviction based on them. This
// refinement is implemented in the un-landed patch at bug 1758274 comment
// 15.
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Main allocation loop complete; "
"beginning spill-bundle allocation loop");
JitSpewCont(JitSpew_RegAlloc,
"\n");
if (!tryAllocatingRegistersForSpillBundles()) {
return false;
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
JitSpew(JitSpew_RegAlloc,
"Spill-bundle allocation loop complete");
JitSpewCont(JitSpew_RegAlloc,
"\n");
// After this point, the VirtualRegister ranges are sorted and must stay
// sorted.
sortVirtualRegisterRanges();
if (!pickStackSlots()) {
return false;
}
#ifdef JS_JITSPEW
if (JitSpewEnabled(JitSpew_RegAlloc)) {
dumpAllocations();
}
#endif
if (!createMoveGroupsFromLiveRangeTransitions()) {
return false;
}
if (!installAllocationsInLIR()) {
return false;
}
if (!populateSafepoints()) {
return false;
}
if (!annotateMoveGroups()) {
return false;
}
JitSpewCont(JitSpew_RegAlloc,
"\n");
if (JitSpewEnabled(JitSpew_RegAlloc)) {
dumpInstructions(
"(Post-allocation LIR)");
}
JitSpew(JitSpew_RegAlloc,
"Finished register allocation");
return true;
}
///////////////////////////////////////////////////////////////////////////////
// //
///////////////////////////////////////////////////////////////////////////////
