10 The vectorization transformation can be rather complicated, involving several
11 potential alternatives, especially for outer-loops [1]_ but also possibly for
12 innermost loops. These alternatives may have significant performance impact,
13 both positive and negative. A cost model is therefore employed to identify the
14 best alternative, including the alternative of avoiding any transformation
17 The Vectorization Plan is an explicit model for describing vectorization
18 candidates. It serves for both optimizing candidates including estimating their
19 cost reliably, and for performing their final translation into IR. This
20 facilitates dealing with multiple vectorization candidates.
24 VPlan is currently used to drive code-generation in LoopVectorize. VPlans are
25 constructed after all cost-based and most legality-related decisions have been
26 taken. As def-use chains between recipes are now fully modeled in VPlan,
27 VPlan-based analyses and transformations are used to simplify and modularize
28 the vectorization process [10]_. Those include transformations to
30 1. Legalize the initial VPlan, e.g. by introducing specialized recipes for
31 reductions and interleave groups.
33 2. Optimize the legalized VPlan, e.g. by removing redundant recipes or
34 introducing active-lane-masks.
36 3. Apply unroll- and vectorization-factor specific optimizations, e.g. removing
37 the backedge to reiterate the vector loop based on VF & UF.
39 Refer to :numref:`fig-vplan-transform-pipeline` for an overview of the current
40 transformation pipeline.
42 Note that some legality checks are already done in VPlan, including checking if
43 all users of a fixed-order recurrence can be re-ordered. This is implemented as
44 a VPlan-to-VPlan transformation that either applies a valid re-ordering or
45 bails out marking the VPlan as invalid.
47 .. _fig-vplan-transform-pipeline:
48 .. figure:: ./vplan-transform-pipeline.png
51 VPlan Transformation Pipeline in 2024
54 VPlan currently models the complete vector loop, as well as additional parts of
55 the vectorization skeleton. Refer to :numref:`fig-vplan-scope` for an overview
56 of the scope covered by VPlan.
59 .. figure:: ./vplan-scope.png
62 Scope modeled in VPlan in 2024
68 Vectorization Workflow
69 ----------------------
70 VPlan-based vectorization involves three major steps, taking a "scenario-based
71 approach" to vectorization planning:
73 1. Legal Step: check if a loop can be legally vectorized; encode constraints and
77 a. Build initial VPlans following the constraints and decisions taken by
78 Legal Step 1, and compute their cost.
79 b. Apply optimizations to the VPlans, possibly forking additional VPlans.
80 Prune sub-optimal VPlans having relatively high cost.
81 3. Execute Step: materialize the best VPlan. Note that this is the only step
86 In what follows, the term "input IR" refers to code that is fed into the
87 vectorizer whereas the term "output IR" refers to code that is generated by the
88 vectorizer. The output IR contains code that has been vectorized or "widened"
89 according to a loop Vectorization Factor (VF), and/or loop unroll-and-jammed
90 according to an Unroll Factor (UF).
91 The design of VPlan follows several high-level guidelines:
93 1. Analysis-like: building and manipulating VPlans must not modify the input IR.
94 In particular, if the best option is not to vectorize at all, the
95 vectorization process terminates before reaching Step 3, and compilation
96 should proceed as if VPlans had not been built.
98 2. Align Cost & Execute: each VPlan must support both estimating the cost and
99 generating the output IR code, such that the cost estimation evaluates the
100 to-be-generated code reliably.
102 3. Support vectorizing additional constructs:
104 a. Outer-loop vectorization. In particular, VPlan must be able to model the
105 control-flow of the output IR which may include multiple basic-blocks and
107 b. SLP vectorization.
108 c. Combinations of the above, including nested vectorization: vectorizing
109 both an inner loop and an outer-loop at the same time (each with its own
110 VF and UF), mixed vectorization: vectorizing a loop with SLP patterns
111 inside [4]_, (re)vectorizing input IR containing vector code.
112 d. Function vectorization [2]_.
114 4. Support multiple candidates efficiently. In particular, similar candidates
115 related to a range of possible VF's and UF's must be represented efficiently.
116 Potential versioning needs to be supported efficiently.
118 5. Support vectorizing idioms, such as interleaved groups of strided loads or
119 stores. This is achieved by modeling a sequence of output instructions using
120 a "Recipe", which is responsible for computing its cost and generating its
123 6. Encapsulate Single-Entry Single-Exit regions (SESE). During vectorization
124 such regions may need to be, for example, predicated and linearized, or
125 replicated VF*UF times to handle scalarized and predicated instructions.
126 Innerloops are also modelled as SESE regions.
128 7. Support instruction-level analysis and transformation, as part of Planning
129 Step 2.b: During vectorization instructions may need to be traversed, moved,
130 replaced by other instructions or be created. For example, vector idiom
131 detection and formation involves searching for and optimizing instruction
136 The low-level design of VPlan comprises of the following classes.
138 :LoopVectorizationPlanner:
139 A LoopVectorizationPlanner is designed to handle the vectorization of a loop
140 or a loop nest. It can construct, optimize and discard one or more VPlans,
141 each VPlan modelling a distinct way to vectorize the loop or the loop nest.
142 Once the best VPlan is determined, including the best VF and UF, this VPlan
143 drives the generation of output IR.
146 A model of a vectorized candidate for a given input IR loop or loop nest. This
147 candidate is represented using a Hierarchical CFG. VPlan supports estimating
148 the cost and driving the generation of the output IR code it represents.
151 A control-flow graph whose nodes are basic-blocks or Hierarchical CFG's. The
152 Hierarchical CFG data structure is similar to the Tile Tree [5]_, where
153 cross-Tile edges are lifted to connect Tiles instead of the original
154 basic-blocks as in Sharir [6]_, promoting the Tile encapsulation. The terms
155 Region and Block are used rather than Tile [5]_ to avoid confusion with loop
159 The building block of the Hierarchical CFG. A pure-virtual base-class of
160 VPBasicBlock and VPRegionBlock, see below. VPBlockBase models the hierarchical
161 control-flow relations with other VPBlocks. Note that in contrast to the IR
162 BasicBlock, a VPBlockBase models its control-flow successors and predecessors
163 directly, rather than through a Terminator branch or through predecessor
164 branches that "use" the VPBlockBase.
167 VPBasicBlock is a subclass of VPBlockBase, and serves as the leaves of the
168 Hierarchical CFG. It represents a sequence of output IR instructions that will
169 appear consecutively in an output IR basic-block. The instructions of this
170 basic-block originate from one or more VPBasicBlocks. VPBasicBlock holds a
171 sequence of zero or more VPRecipes that model the cost and generation of the
172 output IR instructions.
175 VPRegionBlock is a subclass of VPBlockBase. It models a collection of
176 VPBasicBlocks and VPRegionBlocks which form a SESE subgraph of the output IR
177 CFG. A VPRegionBlock may indicate that its contents are to be replicated a
178 constant number of times when output IR is generated, effectively representing
179 a loop with constant trip-count that will be completely unrolled. This is used
180 to support scalarized and predicated instructions with a single model for
181 multiple candidate VF's and UF's.
184 A pure-virtual base class modeling a sequence of one or more output IR
185 instructions, possibly based on one or more input IR instructions. These
186 input IR instructions are referred to as "Ingredients" of the Recipe. A Recipe
187 may specify how its ingredients are to be transformed to produce the output IR
188 instructions; e.g., cloned once, replicated multiple times or widened
189 according to selected VF.
192 The base of VPlan's def-use relations class hierarchy. When instantiated, it
193 models a constant or a live-in Value in VPlan. It has users, which are of type
194 VPUser, but no operands.
197 A VPUser represents an entity that uses a number of VPValues as operands.
198 VPUser is similar in some aspects to LLVM's User class.
201 A VPDef represents an entity that defines zero, one or multiple VPValues.
202 It is used to model the fact that recipes in VPlan can define multiple
206 A VPInstruction is a recipe characterized by a single opcode and optional
207 flags, free of ingredients or other meta-data. VPInstructions also extend
208 LLVM IR's opcodes with idiomatic operations that enrich the Vectorizer's
212 Stores information used for generating output IR, passed from
213 LoopVectorizationPlanner to its selected VPlan for execution, and used to pass
214 additional information down to VPBlocks and VPRecipes.
216 The Planning Process and VPlan Roadmap
217 ======================================
219 Transforming the Loop Vectorizer to use VPlan follows a staged approach. First,
220 VPlan was only used to record the final vectorization decisions, and to execute
221 them: the Hierarchical CFG models the planned control-flow, and Recipes capture
222 decisions taken inside basic-blocks. Currently, VPlan is used also as the basis
223 for taking these decisions, effectively turning them into a series of
224 VPlan-to-VPlan algorithms. Finally, VPlan will support the planning process
225 itself including cost-based analyses for making these decisions, to fully
226 support compositional and iterative decision making.
228 Some decisions are local to an instruction in the loop, such as whether to widen
229 it into a vector instruction or replicate it, keeping the generated instructions
230 in place. Other decisions, however, involve moving instructions, replacing them
231 with other instructions, and/or introducing new instructions. For example, a
232 cast may sink past a later instruction and be widened to handle first-order
233 recurrence; an interleave group of strided gathers or scatters may effectively
234 move to one place where they are replaced with shuffles and a common wide vector
235 load or store; new instructions may be introduced to compute masks, shuffle the
236 elements of vectors, and pack scalar values into vectors or vice-versa.
238 In order for VPlan to support making instruction-level decisions and analyses,
239 it needs to model the relevant instructions along with their def/use relations.
240 This too follows a staged approach: first, the new instructions that compute
241 masks are modeled as VPInstructions, along with their induced def/use subgraph.
242 This effectively models masks in VPlan, facilitating VPlan-based predication.
243 Next, the logic embedded within each Recipe for generating its instructions at
244 VPlan execution time, will instead take part in the planning process by modeling
245 them as VPInstructions. Finally, only logic that applies to instructions as a
246 group will remain in Recipes, such as interleave groups and potentially other
247 idiom groups having synergistic cost.
249 Related LLVM components
250 -----------------------
251 1. SLP Vectorizer: one can compare the VPlan model with LLVM's existing SLP
252 tree, where TSLP [3]_ adds Plan Step 2.b.
254 2. RegionInfo: one can compare VPlan's H-CFG with the Region Analysis as used by
257 3. Loop Vectorizer: the Vectorization Plan aims to upgrade the infrastructure of
258 the Loop Vectorizer and extend it to handle outer loops [8]_, [9]_.
262 .. [1] "Outer-loop vectorization: revisited for short SIMD architectures", Dorit
263 Nuzman and Ayal Zaks, PACT 2008.
265 .. [2] "Proposal for function vectorization and loop vectorization with function
266 calls", Xinmin Tian, [`cfe-dev
267 <http://lists.llvm.org/pipermail/cfe-dev/2016-March/047732.html>`_].,
269 See also `review <https://reviews.llvm.org/D22792>`_.
271 .. [3] "Throttling Automatic Vectorization: When Less is More", Vasileios
272 Porpodas and Tim Jones, PACT 2015 and LLVM Developers' Meeting 2015.
274 .. [4] "Exploiting mixed SIMD parallelism by reducing data reorganization
275 overhead", Hao Zhou and Jingling Xue, CGO 2016.
277 .. [5] "Register Allocation via Hierarchical Graph Coloring", David Callahan and
278 Brian Koblenz, PLDI 1991
280 .. [6] "Structural analysis: A new approach to flow analysis in optimizing
281 compilers", M. Sharir, Journal of Computer Languages, Jan. 1980
283 .. [7] "Enabling Polyhedral Optimizations in LLVM", Tobias Grosser, Diploma
286 .. [8] "Introducing VPlan to the Loop Vectorizer", Gil Rapaport and Ayal Zaks,
287 European LLVM Developers' Meeting 2017.
289 .. [9] "Extending LoopVectorizer: OpenMP4.5 SIMD and Outer Loop
290 Auto-Vectorization", Intel Vectorizer Team, LLVM Developers' Meeting 2016.
292 .. [10] "VPlan: Status Update and Future Roadmap", Ayal Zaks and Florian Hahn,
293 LLVM Developers' Meeting 2023, https://www.youtube.com/watch?v=SzGP4PgMuLE