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.
25 Vectorization Workflow
26 ----------------------
27 VPlan-based vectorization involves three major steps, taking a "scenario-based
28 approach" to vectorization planning:
30 1. Legal Step: check if a loop can be legally vectorized; encode constraints and
34 a. Build initial VPlans following the constraints and decisions taken by
35 Legal Step 1, and compute their cost.
36 b. Apply optimizations to the VPlans, possibly forking additional VPlans.
37 Prune sub-optimal VPlans having relatively high cost.
38 3. Execute Step: materialize the best VPlan. Note that this is the only step
43 In what follows, the term "input IR" refers to code that is fed into the
44 vectorizer whereas the term "output IR" refers to code that is generated by the
45 vectorizer. The output IR contains code that has been vectorized or "widened"
46 according to a loop Vectorization Factor (VF), and/or loop unroll-and-jammed
47 according to an Unroll Factor (UF).
48 The design of VPlan follows several high-level guidelines:
50 1. Analysis-like: building and manipulating VPlans must not modify the input IR.
51 In particular, if the best option is not to vectorize at all, the
52 vectorization process terminates before reaching Step 3, and compilation
53 should proceed as if VPlans had not been built.
55 2. Align Cost & Execute: each VPlan must support both estimating the cost and
56 generating the output IR code, such that the cost estimation evaluates the
57 to-be-generated code reliably.
59 3. Support vectorizing additional constructs:
61 a. Outer-loop vectorization. In particular, VPlan must be able to model the
62 control-flow of the output IR which may include multiple basic-blocks and
65 c. Combinations of the above, including nested vectorization: vectorizing
66 both an inner loop and an outer-loop at the same time (each with its own
67 VF and UF), mixed vectorization: vectorizing a loop with SLP patterns
68 inside [4]_, (re)vectorizing input IR containing vector code.
69 d. Function vectorization [2]_.
71 4. Support multiple candidates efficiently. In particular, similar candidates
72 related to a range of possible VF's and UF's must be represented efficiently.
73 Potential versioning needs to be supported efficiently.
75 5. Support vectorizing idioms, such as interleaved groups of strided loads or
76 stores. This is achieved by modeling a sequence of output instructions using
77 a "Recipe", which is responsible for computing its cost and generating its
80 6. Encapsulate Single-Entry Single-Exit regions (SESE). During vectorization
81 such regions may need to be, for example, predicated and linearized, or
82 replicated VF*UF times to handle scalarized and predicated instructions.
83 Innerloops are also modelled as SESE regions.
85 7. Support instruction-level analysis and transformation, as part of Planning
86 Step 2.b: During vectorization instructions may need to be traversed, moved,
87 replaced by other instructions or be created. For example, vector idiom
88 detection and formation involves searching for and optimizing instruction
93 The low-level design of VPlan comprises of the following classes.
95 :LoopVectorizationPlanner:
96 A LoopVectorizationPlanner is designed to handle the vectorization of a loop
97 or a loop nest. It can construct, optimize and discard one or more VPlans,
98 each VPlan modelling a distinct way to vectorize the loop or the loop nest.
99 Once the best VPlan is determined, including the best VF and UF, this VPlan
100 drives the generation of output IR.
103 A model of a vectorized candidate for a given input IR loop or loop nest. This
104 candidate is represented using a Hierarchical CFG. VPlan supports estimating
105 the cost and driving the generation of the output IR code it represents.
108 A control-flow graph whose nodes are basic-blocks or Hierarchical CFG's. The
109 Hierarchical CFG data structure is similar to the Tile Tree [5]_, where
110 cross-Tile edges are lifted to connect Tiles instead of the original
111 basic-blocks as in Sharir [6]_, promoting the Tile encapsulation. The terms
112 Region and Block are used rather than Tile [5]_ to avoid confusion with loop
116 The building block of the Hierarchical CFG. A pure-virtual base-class of
117 VPBasicBlock and VPRegionBlock, see below. VPBlockBase models the hierarchical
118 control-flow relations with other VPBlocks. Note that in contrast to the IR
119 BasicBlock, a VPBlockBase models its control-flow successors and predecessors
120 directly, rather than through a Terminator branch or through predecessor
121 branches that "use" the VPBlockBase.
124 VPBasicBlock is a subclass of VPBlockBase, and serves as the leaves of the
125 Hierarchical CFG. It represents a sequence of output IR instructions that will
126 appear consecutively in an output IR basic-block. The instructions of this
127 basic-block originate from one or more VPBasicBlocks. VPBasicBlock holds a
128 sequence of zero or more VPRecipes that model the cost and generation of the
129 output IR instructions.
132 VPRegionBlock is a subclass of VPBlockBase. It models a collection of
133 VPBasicBlocks and VPRegionBlocks which form a SESE subgraph of the output IR
134 CFG. A VPRegionBlock may indicate that its contents are to be replicated a
135 constant number of times when output IR is generated, effectively representing
136 a loop with constant trip-count that will be completely unrolled. This is used
137 to support scalarized and predicated instructions with a single model for
138 multiple candidate VF's and UF's.
141 A pure-virtual base class modeling a sequence of one or more output IR
142 instructions, possibly based on one or more input IR instructions. These
143 input IR instructions are referred to as "Ingredients" of the Recipe. A Recipe
144 may specify how its ingredients are to be transformed to produce the output IR
145 instructions; e.g., cloned once, replicated multiple times or widened
146 according to selected VF.
149 The base of VPlan's def-use relations class hierarchy. When instantiated, it
150 models a constant or a live-in Value in VPlan. It has users, which are of type
151 VPUser, but no operands.
154 A VPValue representing a general vertex in the def-use graph of VPlan. It has
155 operands which are of type VPValue. When instantiated, it represents a
156 live-out Instruction that exists outside VPlan. VPUser is similar in some
157 aspects to LLVM's User class.
160 A VPInstruction is both a VPRecipe and a VPUser. It models a single
161 VPlan-level instruction to be generated if the VPlan is executed, including
162 its opcode and possibly additional characteristics. It is the basis for
163 writing instruction-level analyses and optimizations in VPlan as creating,
164 replacing or moving VPInstructions record both def-use and scheduling
165 decisions. VPInstructions also extend LLVM IR's opcodes with idiomatic
166 operations that enrich the Vectorizer's semantics.
169 Stores information used for generating output IR, passed from
170 LoopVectorizationPlanner to its selected VPlan for execution, and used to pass
171 additional information down to VPBlocks and VPRecipes.
173 The Planning Process and VPlan Roadmap
174 ======================================
176 Transforming the Loop Vectorizer to use VPlan follows a staged approach. First,
177 VPlan is used to record the final vectorization decisions, and to execute them:
178 the Hierarchical CFG models the planned control-flow, and Recipes capture
179 decisions taken inside basic-blocks. Next, VPlan will be used also as the basis
180 for taking these decisions, effectively turning them into a series of
181 VPlan-to-VPlan algorithms. Finally, VPlan will support the planning process
182 itself including cost-based analyses for making these decisions, to fully
183 support compositional and iterative decision making.
185 Some decisions are local to an instruction in the loop, such as whether to widen
186 it into a vector instruction or replicate it, keeping the generated instructions
187 in place. Other decisions, however, involve moving instructions, replacing them
188 with other instructions, and/or introducing new instructions. For example, a
189 cast may sink past a later instruction and be widened to handle first-order
190 recurrence; an interleave group of strided gathers or scatters may effectively
191 move to one place where they are replaced with shuffles and a common wide vector
192 load or store; new instructions may be introduced to compute masks, shuffle the
193 elements of vectors, and pack scalar values into vectors or vice-versa.
195 In order for VPlan to support making instruction-level decisions and analyses,
196 it needs to model the relevant instructions along with their def/use relations.
197 This too follows a staged approach: first, the new instructions that compute
198 masks are modeled as VPInstructions, along with their induced def/use subgraph.
199 This effectively models masks in VPlan, facilitating VPlan-based predication.
200 Next, the logic embedded within each Recipe for generating its instructions at
201 VPlan execution time, will instead take part in the planning process by modeling
202 them as VPInstructions. Finally, only logic that applies to instructions as a
203 group will remain in Recipes, such as interleave groups and potentially other
204 idiom groups having synergistic cost.
206 Related LLVM components
207 -----------------------
208 1. SLP Vectorizer: one can compare the VPlan model with LLVM's existing SLP
209 tree, where TSLP [3]_ adds Plan Step 2.b.
211 2. RegionInfo: one can compare VPlan's H-CFG with the Region Analysis as used by
214 3. Loop Vectorizer: the Vectorization Plan aims to upgrade the infrastructure of
215 the Loop Vectorizer and extend it to handle outer loops [8]_, [9]_.
219 .. [1] "Outer-loop vectorization: revisited for short SIMD architectures", Dorit
220 Nuzman and Ayal Zaks, PACT 2008.
222 .. [2] "Proposal for function vectorization and loop vectorization with function
223 calls", Xinmin Tian, [`cfe-dev
224 <http://lists.llvm.org/pipermail/cfe-dev/2016-March/047732.html>`_].,
226 See also `review <https://reviews.llvm.org/D22792>`_.
228 .. [3] "Throttling Automatic Vectorization: When Less is More", Vasileios
229 Porpodas and Tim Jones, PACT 2015 and LLVM Developers' Meeting 2015.
231 .. [4] "Exploiting mixed SIMD parallelism by reducing data reorganization
232 overhead", Hao Zhou and Jingling Xue, CGO 2016.
234 .. [5] "Register Allocation via Hierarchical Graph Coloring", David Callahan and
235 Brian Koblenz, PLDI 1991
237 .. [6] "Structural analysis: A new approach to flow analysis in optimizing
238 compilers", M. Sharir, Journal of Computer Languages, Jan. 1980
240 .. [7] "Enabling Polyhedral Optimizations in LLVM", Tobias Grosser, Diploma
243 .. [8] "Introducing VPlan to the Loop Vectorizer", Gil Rapaport and Ayal Zaks,
244 European LLVM Developers' Meeting 2017.
246 .. [9] "Extending LoopVectorizer: OpenMP4.5 SIMD and Outer Loop
247 Auto-Vectorization", Intel Vectorizer Team, LLVM Developers' Meeting 2016.