xref: /llvm-project/llvm/docs/Frontend/PerformanceTips.rst (revision 1347b9a3aa671d610e812579ab5e5f05870586cf)
1=====================================
2Performance Tips for Frontend Authors
3=====================================
4
5.. contents::
6   :local:
7   :depth: 2
8
9Abstract
10========
11
12The intended audience of this document is developers of language frontends
13targeting LLVM IR. This document is home to a collection of tips on how to
14generate IR that optimizes well.
15
16IR Best Practices
17=================
18
19As with any optimizer, LLVM has its strengths and weaknesses.  In some cases,
20surprisingly small changes in the source IR can have a large effect on the
21generated code.
22
23Beyond the specific items on the list below, it's worth noting that the most
24mature frontend for LLVM is Clang.  As a result, the further your IR gets from
25what Clang might emit, the less likely it is to be effectively optimized. It
26can often be useful to write a quick C program with the semantics you're trying
27to model and see what decisions Clang's IRGen makes about what IR to emit.
28Studying Clang's CodeGen directory can also be a good source of ideas.  Note
29that Clang and LLVM are explicitly version locked so you'll need to make sure
30you're using a Clang built from the same git revision or release as the LLVM
31library you're using.  As always, it's *strongly* recommended that you track
32tip of tree development, particularly during bring up of a new project.
33
34The Basics
35^^^^^^^^^^^
36
37#. Make sure that your Modules contain both a data layout specification and
38   target triple. Without these pieces, non of the target specific optimization
39   will be enabled.  This can have a major effect on the generated code quality.
40
41#. For each function or global emitted, use the most private linkage type
42   possible (private, internal or linkonce_odr preferably).  Doing so will
43   make LLVM's inter-procedural optimizations much more effective.
44
45#. Avoid high in-degree basic blocks (e.g. basic blocks with dozens or hundreds
46   of predecessors).  Among other issues, the register allocator is known to
47   perform badly with confronted with such structures.  The only exception to
48   this guidance is that a unified return block with high in-degree is fine.
49
50Use of allocas
51^^^^^^^^^^^^^^
52
53An alloca instruction can be used to represent a function scoped stack slot,
54but can also represent dynamic frame expansion.  When representing function
55scoped variables or locations, placing alloca instructions at the beginning of
56the entry block should be preferred.   In particular, place them before any
57call instructions. Call instructions might get inlined and replaced with
58multiple basic blocks. The end result is that a following alloca instruction
59would no longer be in the entry basic block afterward.
60
61The SROA (Scalar Replacement Of Aggregates) and Mem2Reg passes only attempt
62to eliminate alloca instructions that are in the entry basic block.  Given
63SSA is the canonical form expected by much of the optimizer; if allocas can
64not be eliminated by Mem2Reg or SROA, the optimizer is likely to be less
65effective than it could be.
66
67Avoid creating values of aggregate type
68^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
69
70Avoid creating values of :ref:`aggregate types <t_aggregate>` (i.e. structs and
71arrays). In particular, avoid loading and storing them, or manipulating them
72with insertvalue and extractvalue instructions. Instead, only load and store
73individual fields of the aggregate.
74
75There are some exceptions to this rule:
76
77* It is fine to use values of aggregate type in global variable initializers.
78* It is fine to return structs, if this is done to represent the return of
79  multiple values in registers.
80* It is fine to work with structs returned by LLVM intrinsics, such as the
81  ``with.overflow`` family of intrinsics.
82* It is fine to use aggregate *types* without creating values. For example,
83  they are commonly used in ``getelementptr`` instructions or attributes like
84  ``sret``.
85
86Avoid loads and stores of non-byte-sized types
87^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
88
89Avoid loading or storing non-byte-sized types like ``i1``. Instead,
90appropriately extend them to the next byte-sized type.
91
92For example, when working with boolean values, store them by zero-extending
93``i1`` to ``i8`` and load them by loading ``i8`` and truncating to ``i1``.
94
95If you do use loads/stores on non-byte-sized types, make sure that you *always*
96use those types. For example, do not first store ``i8`` and then load ``i1``.
97
98Prefer zext over sext when legal
99^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
100
101On some architectures (X86_64 is one), sign extension can involve an extra
102instruction whereas zero extension can be folded into a load.  LLVM will try to
103replace a sext with a zext when it can be proven safe, but if you have
104information in your source language about the range of an integer value, it can
105be profitable to use a zext rather than a sext.
106
107Alternatively, you can :ref:`specify the range of the value using metadata
108<range-metadata>` and LLVM can do the sext to zext conversion for you.
109
110Zext GEP indices to machine register width
111^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
112
113Internally, LLVM often promotes the width of GEP indices to machine register
114width.  When it does so, it will default to using sign extension (sext)
115operations for safety.  If your source language provides information about
116the range of the index, you may wish to manually extend indices to machine
117register width using a zext instruction.
118
119When to specify alignment
120^^^^^^^^^^^^^^^^^^^^^^^^^^
121LLVM will always generate correct code if you don’t specify alignment, but may
122generate inefficient code.  For example, if you are targeting MIPS (or older
123ARM ISAs) then the hardware does not handle unaligned loads and stores, and
124so you will enter a trap-and-emulate path if you do a load or store with
125lower-than-natural alignment.  To avoid this, LLVM will emit a slower
126sequence of loads, shifts and masks (or load-right + load-left on MIPS) for
127all cases where the load / store does not have a sufficiently high alignment
128in the IR.
129
130The alignment is used to guarantee the alignment on allocas and globals,
131though in most cases this is unnecessary (most targets have a sufficiently
132high default alignment that they’ll be fine).  It is also used to provide a
133contract to the back end saying ‘either this load/store has this alignment, or
134it is undefined behavior’.  This means that the back end is free to emit
135instructions that rely on that alignment (and mid-level optimizers are free to
136perform transforms that require that alignment).  For x86, it doesn’t make
137much difference, as almost all instructions are alignment-independent.  For
138MIPS, it can make a big difference.
139
140Note that if your loads and stores are atomic, the backend will be unable to
141lower an under aligned access into a sequence of natively aligned accesses.
142As a result, alignment is mandatory for atomic loads and stores.
143
144Other Things to Consider
145^^^^^^^^^^^^^^^^^^^^^^^^
146
147#. Use ptrtoint/inttoptr sparingly (they interfere with pointer aliasing
148   analysis), prefer GEPs
149
150#. Prefer globals over inttoptr of a constant address - this gives you
151   dereferencability information.  In MCJIT, use getSymbolAddress to provide
152   actual address.
153
154#. Be wary of ordered and atomic memory operations.  They are hard to optimize
155   and may not be well optimized by the current optimizer.  Depending on your
156   source language, you may consider using fences instead.
157
158#. If calling a function which is known to throw an exception (unwind), use
159   an invoke with a normal destination which contains an unreachable
160   instruction.  This form conveys to the optimizer that the call returns
161   abnormally.  For an invoke which neither returns normally or requires unwind
162   code in the current function, you can use a noreturn call instruction if
163   desired.  This is generally not required because the optimizer will convert
164   an invoke with an unreachable unwind destination to a call instruction.
165
166#. Use profile metadata to indicate statically known cold paths, even if
167   dynamic profiling information is not available.  This can make a large
168   difference in code placement and thus the performance of tight loops.
169
170#. When generating code for loops, try to avoid terminating the header block of
171   the loop earlier than necessary.  If the terminator of the loop header
172   block is a loop exiting conditional branch, the effectiveness of LICM will
173   be limited for loads not in the header.  (This is due to the fact that LLVM
174   may not know such a load is safe to speculatively execute and thus can't
175   lift an otherwise loop invariant load unless it can prove the exiting
176   condition is not taken.)  It can be profitable, in some cases, to emit such
177   instructions into the header even if they are not used along a rarely
178   executed path that exits the loop.  This guidance specifically does not
179   apply if the condition which terminates the loop header is itself invariant,
180   or can be easily discharged by inspecting the loop index variables.
181
182#. In hot loops, consider duplicating instructions from small basic blocks
183   which end in highly predictable terminators into their successor blocks.
184   If a hot successor block contains instructions which can be vectorized
185   with the duplicated ones, this can provide a noticeable throughput
186   improvement.  Note that this is not always profitable and does involve a
187   potentially large increase in code size.
188
189#. When checking a value against a constant, emit the check using a consistent
190   comparison type.  The GVN pass *will* optimize redundant equalities even if
191   the type of comparison is inverted, but GVN only runs late in the pipeline.
192   As a result, you may miss the opportunity to run other important
193   optimizations.
194
195#. Avoid using arithmetic intrinsics unless you are *required* by your source
196   language specification to emit a particular code sequence.  The optimizer
197   is quite good at reasoning about general control flow and arithmetic, it is
198   not anywhere near as strong at reasoning about the various intrinsics.  If
199   profitable for code generation purposes, the optimizer will likely form the
200   intrinsics itself late in the optimization pipeline.  It is *very* rarely
201   profitable to emit these directly in the language frontend.  This item
202   explicitly includes the use of the :ref:`overflow intrinsics <int_overflow>`.
203
204#. Avoid using the :ref:`assume intrinsic <int_assume>` until you've
205   established that a) there's no other way to express the given fact and b)
206   that fact is critical for optimization purposes.  Assumes are a great
207   prototyping mechanism, but they can have negative effects on both compile
208   time and optimization effectiveness.  The former is fixable with enough
209   effort, but the later is fairly fundamental to their designed purpose.  If
210   you are creating a non-terminator unreachable instruction or passing a false
211   value, use the ``store i1 true, ptr poison, align 1`` canonical form.
212
213
214Describing Language Specific Properties
215=======================================
216
217When translating a source language to LLVM, finding ways to express concepts
218and guarantees available in your source language which are not natively
219provided by LLVM IR will greatly improve LLVM's ability to optimize your code.
220As an example, C/C++'s ability to mark every add as "no signed wrap (nsw)" goes
221a long way to assisting the optimizer in reasoning about loop induction
222variables and thus generating more optimal code for loops.
223
224The LLVM LangRef includes a number of mechanisms for annotating the IR with
225additional semantic information.  It is *strongly* recommended that you become
226highly familiar with this document.  The list below is intended to highlight a
227couple of items of particular interest, but is by no means exhaustive.
228
229Restricted Operation Semantics
230^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
231#. Add nsw/nuw flags as appropriate.  Reasoning about overflow is
232   generally hard for an optimizer so providing these facts from the frontend
233   can be very impactful.
234
235#. Use fast-math flags on floating point operations if legal.  If you don't
236   need strict IEEE floating point semantics, there are a number of additional
237   optimizations that can be performed.  This can be highly impactful for
238   floating point intensive computations.
239
240Describing Aliasing Properties
241^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
242
243#. Add noalias/align/dereferenceable/nonnull to function arguments and return
244   values as appropriate
245
246#. Use pointer aliasing metadata, especially tbaa metadata, to communicate
247   otherwise-non-deducible pointer aliasing facts
248
249#. Use inbounds on geps.  This can help to disambiguate some aliasing queries.
250
251Undefined Values
252^^^^^^^^^^^^^^^^
253
254#. Use poison values instead of undef values whenever possible.
255
256#. Tag function parameters with the noundef attribute whenever possible.
257
258Modeling Memory Effects
259^^^^^^^^^^^^^^^^^^^^^^^^
260
261#. Mark functions as readnone/readonly/argmemonly or noreturn/nounwind when
262   known.  The optimizer will try to infer these flags, but may not always be
263   able to.  Manual annotations are particularly important for external
264   functions that the optimizer can not analyze.
265
266#. Use the lifetime.start/lifetime.end and invariant.start/invariant.end
267   intrinsics where possible.  Common profitable uses are for stack like data
268   structures (thus allowing dead store elimination) and for describing
269   life times of allocas (thus allowing smaller stack sizes).
270
271#. Mark invariant locations using !invariant.load and TBAA's constant flags
272
273Pass Ordering
274^^^^^^^^^^^^^
275
276One of the most common mistakes made by new language frontend projects is to
277use the existing -O2 or -O3 pass pipelines as is.  These pass pipelines make a
278good starting point for an optimizing compiler for any language, but they have
279been carefully tuned for C and C++, not your target language.  You will almost
280certainly need to use a custom pass order to achieve optimal performance.  A
281couple specific suggestions:
282
283#. For languages with numerous rarely executed guard conditions (e.g. null
284   checks, type checks, range checks) consider adding an extra execution or
285   two of LoopUnswitch and LICM to your pass order.  The standard pass order,
286   which is tuned for C and C++ applications, may not be sufficient to remove
287   all dischargeable checks from loops.
288
289#. If your language uses range checks, consider using the IRCE pass.  It is not
290   currently part of the standard pass order.
291
292#. A useful sanity check to run is to run your optimized IR back through the
293   -O2 pipeline again.  If you see noticeable improvement in the resulting IR,
294   you likely need to adjust your pass order.
295
296
297I Still Can't Find What I'm Looking For
298^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
299
300If you didn't find what you were looking for above, consider proposing a piece
301of metadata which provides the optimization hint you need.  Such extensions are
302relatively common and are generally well received by the community.  You will
303need to ensure that your proposal is sufficiently general so that it benefits
304others if you wish to contribute it upstream.
305
306You should also consider describing the problem you're facing on `Discourse
307<https://discourse.llvm.org>`_ and asking for advice.
308It's entirely possible someone has encountered your problem before and can
309give good advice.  If there are multiple interested parties, that also
310increases the chances that a metadata extension would be well received by the
311community as a whole.
312
313Adding to this document
314=======================
315
316If you run across a case that you feel deserves to be covered here, please send
317a patch to `llvm-commits
318<http://lists.llvm.org/mailman/listinfo/llvm-commits>`_ for review.
319
320If you have questions on these items, please ask them on `Discourse
321<https://discourse.llvm.org>`_.  The more relevant
322context you are able to give to your question, the more likely it is to be
323answered.
324