• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1=========
2MemorySSA
3=========
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11``MemorySSA`` is an analysis that allows us to cheaply reason about the
12interactions between various memory operations. Its goal is to replace
13``MemoryDependenceAnalysis`` for most (if not all) use-cases. This is because,
14unless you're very careful, use of ``MemoryDependenceAnalysis`` can easily
15result in quadratic-time algorithms in LLVM. Additionally, ``MemorySSA`` doesn't
16have as many arbitrary limits as ``MemoryDependenceAnalysis``, so you should get
17better results, too. One common use of ``MemorySSA`` is to quickly find out
18that something definitely cannot happen (for example, reason that a hoist
19out of a loop can't happen).
20
21At a high level, one of the goals of ``MemorySSA`` is to provide an SSA based
22form for memory, complete with def-use and use-def chains, which
23enables users to quickly find may-def and may-uses of memory operations.
24It can also be thought of as a way to cheaply give versions to the complete
25state of memory, and associate memory operations with those versions.
26
27This document goes over how ``MemorySSA`` is structured, and some basic
28intuition on how ``MemorySSA`` works.
29
30A paper on MemorySSA (with notes about how it's implemented in GCC) `can be
31found here <http://www.airs.com/dnovillo/Papers/mem-ssa.pdf>`_. Though, it's
32relatively out-of-date; the paper references multiple memory partitions, but GCC
33eventually swapped to just using one, like we now have in LLVM.  Like
34GCC's, LLVM's MemorySSA is intraprocedural.
35
36
37MemorySSA Structure
38===================
39
40MemorySSA is a virtual IR. After it's built, ``MemorySSA`` will contain a
41structure that maps ``Instruction``\ s to ``MemoryAccess``\ es, which are
42``MemorySSA``'s parallel to LLVM ``Instruction``\ s.
43
44Each ``MemoryAccess`` can be one of three types:
45
46- ``MemoryDef``
47- ``MemoryPhi``
48- ``MemoryUse``
49
50``MemoryDef``\ s are operations which may either modify memory, or which
51introduce some kind of ordering constraints. Examples of ``MemoryDef``\ s
52include ``store``\ s, function calls, ``load``\ s with ``acquire`` (or higher)
53ordering, volatile operations, memory fences, etc. A ``MemoryDef``
54always introduces a new version of the entire memory and is linked with a single
55``MemoryDef/MemoryPhi`` which is the version of memory that the new
56version is based on. This implies that there is a *single*
57``Def`` chain that connects all the ``Def``\ s, either directly
58or indirectly. For example in:
59
60.. code-block:: llvm
61
62  b = MemoryDef(a)
63  c = MemoryDef(b)
64  d = MemoryDef(c)
65
66``d`` is connected directly with ``c`` and indirectly with ``b``.
67This means that ``d`` potentially clobbers (see below) ``c`` *or*
68``b`` *or* both. This in turn implies that without the use of `The walker`_,
69initially every ``MemoryDef`` clobbers every other ``MemoryDef``.
70
71``MemoryPhi``\ s are ``PhiNode``\ s, but for memory operations. If at any
72point we have two (or more) ``MemoryDef``\ s that could flow into a
73``BasicBlock``, the block's top ``MemoryAccess`` will be a
74``MemoryPhi``. As in LLVM IR, ``MemoryPhi``\ s don't correspond to any
75concrete operation. As such, ``BasicBlock``\ s are mapped to ``MemoryPhi``\ s
76inside ``MemorySSA``, whereas ``Instruction``\ s are mapped to ``MemoryUse``\ s
77and ``MemoryDef``\ s.
78
79Note also that in SSA, Phi nodes merge must-reach definitions (that is,
80definitions that *must* be new versions of variables). In MemorySSA, PHI nodes
81merge may-reach definitions (that is, until disambiguated, the versions that
82reach a phi node may or may not clobber a given variable).
83
84``MemoryUse``\ s are operations which use but don't modify memory. An example of
85a ``MemoryUse`` is a ``load``, or a ``readonly`` function call.
86
87Every function that exists has a special ``MemoryDef`` called ``liveOnEntry``.
88It dominates every ``MemoryAccess`` in the function that ``MemorySSA`` is being
89run on, and implies that we've hit the top of the function. It's the only
90``MemoryDef`` that maps to no ``Instruction`` in LLVM IR. Use of
91``liveOnEntry`` implies that the memory being used is either undefined or
92defined before the function begins.
93
94An example of all of this overlaid on LLVM IR (obtained by running ``opt
95-passes='print<memoryssa>' -disable-output`` on an ``.ll`` file) is below. When
96viewing this example, it may be helpful to view it in terms of clobbers.
97The operands of a given ``MemoryAccess`` are all (potential) clobbers of said
98``MemoryAccess``, and the value produced by a ``MemoryAccess`` can act as a clobber
99for other ``MemoryAccess``\ es.
100
101If a ``MemoryAccess`` is a *clobber* of another, it means that these two
102``MemoryAccess``\ es may access the same memory. For example, ``x = MemoryDef(y)``
103means that ``x`` potentially modifies memory that ``y`` modifies/constrains
104(or has modified / constrained).
105In the same manner, ``a = MemoryPhi({BB1,b},{BB2,c})`` means that
106anyone that uses ``a`` is accessing memory potentially modified / constrained
107by either ``b`` or ``c`` (or both).  And finally, ``MemoryUse(x)`` means
108that this use accesses memory that ``x`` has modified / constrained
109(as an example, think that if ``x = MemoryDef(...)``
110and ``MemoryUse(x)`` are in the same loop, the use can't
111be hoisted outside alone).
112
113Another useful way of looking at it is in terms of memory versions.
114In that view, operands of a given ``MemoryAccess`` are the version
115of the entire memory before the operation, and if the access produces
116a value (i.e. ``MemoryDef/MemoryPhi``),
117the value is the new version of the memory after the operation.
118
119.. code-block:: llvm
120
121  define void @foo() {
122  entry:
123    %p1 = alloca i8
124    %p2 = alloca i8
125    %p3 = alloca i8
126    ; 1 = MemoryDef(liveOnEntry)
127    store i8 0, i8* %p3
128    br label %while.cond
129
130  while.cond:
131    ; 6 = MemoryPhi({entry,1},{if.end,4})
132    br i1 undef, label %if.then, label %if.else
133
134  if.then:
135    ; 2 = MemoryDef(6)
136    store i8 0, i8* %p1
137    br label %if.end
138
139  if.else:
140    ; 3 = MemoryDef(6)
141    store i8 1, i8* %p2
142    br label %if.end
143
144  if.end:
145    ; 5 = MemoryPhi({if.then,2},{if.else,3})
146    ; MemoryUse(5)
147    %1 = load i8, i8* %p1
148    ; 4 = MemoryDef(5)
149    store i8 2, i8* %p2
150    ; MemoryUse(1)
151    %2 = load i8, i8* %p3
152    br label %while.cond
153  }
154
155The ``MemorySSA`` IR is shown in comments that precede the instructions they map
156to (if such an instruction exists). For example, ``1 = MemoryDef(liveOnEntry)``
157is a ``MemoryAccess`` (specifically, a ``MemoryDef``), and it describes the LLVM
158instruction ``store i8 0, i8* %p3``. Other places in ``MemorySSA`` refer to this
159particular ``MemoryDef`` as ``1`` (much like how one can refer to ``load i8, i8*
160%p1`` in LLVM with ``%1``). Again, ``MemoryPhi``\ s don't correspond to any LLVM
161Instruction, so the line directly below a ``MemoryPhi`` isn't special.
162
163Going from the top down:
164
165- ``6 = MemoryPhi({entry,1},{if.end,4})`` notes that, when entering
166  ``while.cond``, the reaching definition for it is either ``1`` or ``4``. This
167  ``MemoryPhi`` is referred to in the textual IR by the number ``6``.
168- ``2 = MemoryDef(6)`` notes that ``store i8 0, i8* %p1`` is a definition,
169  and its reaching definition before it is ``6``, or the ``MemoryPhi`` after
170  ``while.cond``. (See the `Build-time use optimization`_ and `Precision`_
171  sections below for why this ``MemoryDef`` isn't linked to a separate,
172  disambiguated ``MemoryPhi``.)
173- ``3 = MemoryDef(6)`` notes that ``store i8 0, i8* %p2`` is a definition; its
174  reaching definition is also ``6``.
175- ``5 = MemoryPhi({if.then,2},{if.else,3})`` notes that the clobber before
176  this block could either be ``2`` or ``3``.
177- ``MemoryUse(5)`` notes that ``load i8, i8* %p1`` is a use of memory, and that
178  it's clobbered by ``5``.
179- ``4 = MemoryDef(5)`` notes that ``store i8 2, i8* %p2`` is a definition; it's
180  reaching definition is ``5``.
181- ``MemoryUse(1)`` notes that ``load i8, i8* %p3`` is just a user of memory,
182  and the last thing that could clobber this use is above ``while.cond`` (e.g.
183  the store to ``%p3``). In memory versioning parlance, it really only depends on
184  the memory version 1, and is unaffected by the new memory versions generated since
185  then.
186
187As an aside, ``MemoryAccess`` is a ``Value`` mostly for convenience; it's not
188meant to interact with LLVM IR.
189
190Design of MemorySSA
191===================
192
193``MemorySSA`` is an analysis that can be built for any arbitrary function. When
194it's built, it does a pass over the function's IR in order to build up its
195mapping of ``MemoryAccess``\ es. You can then query ``MemorySSA`` for things
196like the dominance relation between ``MemoryAccess``\ es, and get the
197``MemoryAccess`` for any given ``Instruction`` .
198
199When ``MemorySSA`` is done building, it also hands you a ``MemorySSAWalker``
200that you can use (see below).
201
202
203The walker
204----------
205
206A structure that helps ``MemorySSA`` do its job is the ``MemorySSAWalker``, or
207the walker, for short. The goal of the walker is to provide answers to clobber
208queries beyond what's represented directly by ``MemoryAccess``\ es. For example,
209given:
210
211.. code-block:: llvm
212
213  define void @foo() {
214    %a = alloca i8
215    %b = alloca i8
216
217    ; 1 = MemoryDef(liveOnEntry)
218    store i8 0, i8* %a
219    ; 2 = MemoryDef(1)
220    store i8 0, i8* %b
221  }
222
223The store to ``%a`` is clearly not a clobber for the store to ``%b``. It would
224be the walker's goal to figure this out, and return ``liveOnEntry`` when queried
225for the clobber of ``MemoryAccess`` ``2``.
226
227By default, ``MemorySSA`` provides a walker that can optimize ``MemoryDef``\ s
228and ``MemoryUse``\ s by consulting whatever alias analysis stack you happen to
229be using. Walkers were built to be flexible, though, so it's entirely reasonable
230(and expected) to create more specialized walkers (e.g. one that specifically
231queries ``GlobalsAA``, one that always stops at ``MemoryPhi`` nodes, etc).
232
233
234Locating clobbers yourself
235^^^^^^^^^^^^^^^^^^^^^^^^^^
236
237If you choose to make your own walker, you can find the clobber for a
238``MemoryAccess`` by walking every ``MemoryDef`` that dominates said
239``MemoryAccess``. The structure of ``MemoryDef``\ s makes this relatively simple;
240they ultimately form a linked list of every clobber that dominates the
241``MemoryAccess`` that you're trying to optimize. In other words, the
242``definingAccess`` of a ``MemoryDef`` is always the nearest dominating
243``MemoryDef`` or ``MemoryPhi`` of said ``MemoryDef``.
244
245
246Build-time use optimization
247---------------------------
248
249``MemorySSA`` will optimize some ``MemoryAccess``\ es at build-time.
250Specifically, we optimize the operand of every ``MemoryUse`` to point to the
251actual clobber of said ``MemoryUse``. This can be seen in the above example; the
252second ``MemoryUse`` in ``if.end`` has an operand of ``1``, which is a
253``MemoryDef`` from the entry block.  This is done to make walking,
254value numbering, etc, faster and easier.
255
256It is not possible to optimize ``MemoryDef`` in the same way, as we
257restrict ``MemorySSA`` to one memory variable and, thus, one Phi node
258per block.
259
260
261Invalidation and updating
262-------------------------
263
264Because ``MemorySSA`` keeps track of LLVM IR, it needs to be updated whenever
265the IR is updated. "Update", in this case, includes the addition, deletion, and
266motion of ``Instructions``. The update API is being made on an as-needed basis.
267If you'd like examples, ``GVNHoist`` is a user of ``MemorySSA``\ s update API.
268
269
270Phi placement
271^^^^^^^^^^^^^
272
273``MemorySSA`` only places ``MemoryPhi``\ s where they're actually
274needed. That is, it is a pruned SSA form, like LLVM's SSA form.  For
275example, consider:
276
277.. code-block:: llvm
278
279  define void @foo() {
280  entry:
281    %p1 = alloca i8
282    %p2 = alloca i8
283    %p3 = alloca i8
284    ; 1 = MemoryDef(liveOnEntry)
285    store i8 0, i8* %p3
286    br label %while.cond
287
288  while.cond:
289    ; 3 = MemoryPhi({%0,1},{if.end,2})
290    br i1 undef, label %if.then, label %if.else
291
292  if.then:
293    br label %if.end
294
295  if.else:
296    br label %if.end
297
298  if.end:
299    ; MemoryUse(1)
300    %1 = load i8, i8* %p1
301    ; 2 = MemoryDef(3)
302    store i8 2, i8* %p2
303    ; MemoryUse(1)
304    %2 = load i8, i8* %p3
305    br label %while.cond
306  }
307
308Because we removed the stores from ``if.then`` and ``if.else``, a ``MemoryPhi``
309for ``if.end`` would be pointless, so we don't place one. So, if you need to
310place a ``MemoryDef`` in ``if.then`` or ``if.else``, you'll need to also create
311a ``MemoryPhi`` for ``if.end``.
312
313If it turns out that this is a large burden, we can just place ``MemoryPhi``\ s
314everywhere. Because we have Walkers that are capable of optimizing above said
315phis, doing so shouldn't prohibit optimizations.
316
317
318Non-Goals
319---------
320
321``MemorySSA`` is meant to reason about the relation between memory
322operations, and enable quicker querying.
323It isn't meant to be the single source of truth for all potential memory-related
324optimizations. Specifically, care must be taken when trying to use ``MemorySSA``
325to reason about atomic or volatile operations, as in:
326
327.. code-block:: llvm
328
329  define i8 @foo(i8* %a) {
330  entry:
331    br i1 undef, label %if.then, label %if.end
332
333  if.then:
334    ; 1 = MemoryDef(liveOnEntry)
335    %0 = load volatile i8, i8* %a
336    br label %if.end
337
338  if.end:
339    %av = phi i8 [0, %entry], [%0, %if.then]
340    ret i8 %av
341  }
342
343Going solely by ``MemorySSA``'s analysis, hoisting the ``load`` to ``entry`` may
344seem legal. Because it's a volatile load, though, it's not.
345
346
347Design tradeoffs
348----------------
349
350Precision
351^^^^^^^^^
352
353``MemorySSA`` in LLVM deliberately trades off precision for speed.
354Let us think about memory variables as if they were disjoint partitions of the
355memory (that is, if you have one variable, as above, it represents the entire
356memory, and if you have multiple variables, each one represents some
357disjoint portion of the memory)
358
359First, because alias analysis results conflict with each other, and
360each result may be what an analysis wants (IE
361TBAA may say no-alias, and something else may say must-alias), it is
362not possible to partition the memory the way every optimization wants.
363Second, some alias analysis results are not transitive (IE A noalias B,
364and B noalias C, does not mean A noalias C), so it is not possible to
365come up with a precise partitioning in all cases without variables to
366represent every pair of possible aliases.  Thus, partitioning
367precisely may require introducing at least N^2 new virtual variables,
368phi nodes, etc.
369
370Each of these variables may be clobbered at multiple def sites.
371
372To give an example, if you were to split up struct fields into
373individual variables, all aliasing operations that may-def multiple struct
374fields, will may-def more than one of them.  This is pretty common (calls,
375copies, field stores, etc).
376
377Experience with SSA forms for memory in other compilers has shown that
378it is simply not possible to do this precisely, and in fact, doing it
379precisely is not worth it, because now all the optimizations have to
380walk tons and tons of virtual variables and phi nodes.
381
382So we partition.  At the point at which you partition, again,
383experience has shown us there is no point in partitioning to more than
384one variable.  It simply generates more IR, and optimizations still
385have to query something to disambiguate further anyway.
386
387As a result, LLVM partitions to one variable.
388
389Use Optimization
390^^^^^^^^^^^^^^^^
391
392Unlike other partitioned forms, LLVM's ``MemorySSA`` does make one
393useful guarantee - all loads are optimized to point at the thing that
394actually clobbers them. This gives some nice properties.  For example,
395for a given store, you can find all loads actually clobbered by that
396store by walking the immediate uses of the store.
397