Home | History | Annotate | Download | only in Frontend
      1 =====================================
      2 Performance Tips for Frontend Authors
      3 =====================================
      4 
      5 .. contents::
      6    :local:
      7    :depth: 2
      8 
      9 Abstract
     10 ========
     11 
     12 The intended audience of this document is developers of language frontends 
     13 targeting LLVM IR. This document is home to a collection of tips on how to 
     14 generate IR that optimizes well.  
     15 
     16 IR Best Practices
     17 =================
     18 
     19 As with any optimizer, LLVM has its strengths and weaknesses.  In some cases, 
     20 surprisingly small changes in the source IR can have a large effect on the 
     21 generated code.  
     22 
     23 Beyond the specific items on the list below, it's worth noting that the most 
     24 mature frontend for LLVM is Clang.  As a result, the further your IR gets from what Clang might emit, the less likely it is to be effectively optimized.  It 
     25 can often be useful to write a quick C program with the semantics you're trying
     26 to model and see what decisions Clang's IRGen makes about what IR to emit.  
     27 Studying Clang's CodeGen directory can also be a good source of ideas.  Note 
     28 that Clang and LLVM are explicitly version locked so you'll need to make sure 
     29 you're using a Clang built from the same svn revision or release as the LLVM 
     30 library you're using.  As always, it's *strongly* recommended that you track 
     31 tip of tree development, particularly during bring up of a new project.
     32 
     33 The Basics
     34 ^^^^^^^^^^^
     35 
     36 #. Make sure that your Modules contain both a data layout specification and 
     37    target triple. Without these pieces, non of the target specific optimization
     38    will be enabled.  This can have a major effect on the generated code quality.
     39 
     40 #. For each function or global emitted, use the most private linkage type
     41    possible (private, internal or linkonce_odr preferably).  Doing so will 
     42    make LLVM's inter-procedural optimizations much more effective.
     43 
     44 #. Avoid high in-degree basic blocks (e.g. basic blocks with dozens or hundreds
     45    of predecessors).  Among other issues, the register allocator is known to 
     46    perform badly with confronted with such structures.  The only exception to 
     47    this guidance is that a unified return block with high in-degree is fine.
     48 
     49 Use of allocas
     50 ^^^^^^^^^^^^^^
     51 
     52 An alloca instruction can be used to represent a function scoped stack slot, 
     53 but can also represent dynamic frame expansion.  When representing function 
     54 scoped variables or locations, placing alloca instructions at the beginning of 
     55 the entry block should be preferred.   In particular, place them before any 
     56 call instructions. Call instructions might get inlined and replaced with 
     57 multiple basic blocks. The end result is that a following alloca instruction 
     58 would no longer be in the entry basic block afterward.
     59 
     60 The SROA (Scalar Replacement Of Aggregates) and Mem2Reg passes only attempt
     61 to eliminate alloca instructions that are in the entry basic block.  Given 
     62 SSA is the canonical form expected by much of the optimizer; if allocas can 
     63 not be eliminated by Mem2Reg or SROA, the optimizer is likely to be less 
     64 effective than it could be.
     65 
     66 Avoid loads and stores of large aggregate type
     67 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
     68 
     69 LLVM currently does not optimize well loads and stores of large :ref:`aggregate
     70 types <t_aggregate>` (i.e. structs and arrays).  As an alternative, consider 
     71 loading individual fields from memory.
     72 
     73 Aggregates that are smaller than the largest (performant) load or store 
     74 instruction supported by the targeted hardware are well supported.  These can 
     75 be an effective way to represent collections of small packed fields.  
     76 
     77 Prefer zext over sext when legal
     78 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
     79 
     80 On some architectures (X86_64 is one), sign extension can involve an extra 
     81 instruction whereas zero extension can be folded into a load.  LLVM will try to
     82 replace a sext with a zext when it can be proven safe, but if you have 
     83 information in your source language about the range of a integer value, it can 
     84 be profitable to use a zext rather than a sext.  
     85 
     86 Alternatively, you can :ref:`specify the range of the value using metadata 
     87 <range-metadata>` and LLVM can do the sext to zext conversion for you.
     88 
     89 Zext GEP indices to machine register width
     90 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
     91 
     92 Internally, LLVM often promotes the width of GEP indices to machine register
     93 width.  When it does so, it will default to using sign extension (sext) 
     94 operations for safety.  If your source language provides information about 
     95 the range of the index, you may wish to manually extend indices to machine 
     96 register width using a zext instruction.
     97 
     98 When to specify alignment
     99 ^^^^^^^^^^^^^^^^^^^^^^^^^^
    100 LLVM will always generate correct code if you dont specify alignment, but may
    101 generate inefficient code.  For example, if you are targeting MIPS (or older 
    102 ARM ISAs) then the hardware does not handle unaligned loads and stores, and 
    103 so you will enter a trap-and-emulate path if you do a load or store with 
    104 lower-than-natural alignment.  To avoid this, LLVM will emit a slower 
    105 sequence of loads, shifts and masks (or load-right + load-left on MIPS) for 
    106 all cases where the load / store does not have a sufficiently high alignment 
    107 in the IR.
    108 
    109 The alignment is used to guarantee the alignment on allocas and globals, 
    110 though in most cases this is unnecessary (most targets have a sufficiently 
    111 high default alignment that theyll be fine).  It is also used to provide a 
    112 contract to the back end saying either this load/store has this alignment, or
    113 it is undefined behavior.  This means that the back end is free to emit 
    114 instructions that rely on that alignment (and mid-level optimizers are free to 
    115 perform transforms that require that alignment).  For x86, it doesnt make 
    116 much difference, as almost all instructions are alignment-independent.  For 
    117 MIPS, it can make a big difference.
    118 
    119 Note that if your loads and stores are atomic, the backend will be unable to 
    120 lower an under aligned access into a sequence of natively aligned accesses.  
    121 As a result, alignment is mandatory for atomic loads and stores.
    122 
    123 Other Things to Consider
    124 ^^^^^^^^^^^^^^^^^^^^^^^^
    125 
    126 #. Use ptrtoint/inttoptr sparingly (they interfere with pointer aliasing 
    127    analysis), prefer GEPs
    128 
    129 #. Prefer globals over inttoptr of a constant address - this gives you 
    130    dereferencability information.  In MCJIT, use getSymbolAddress to provide 
    131    actual address.
    132 
    133 #. Be wary of ordered and atomic memory operations.  They are hard to optimize 
    134    and may not be well optimized by the current optimizer.  Depending on your
    135    source language, you may consider using fences instead.
    136 
    137 #. If calling a function which is known to throw an exception (unwind), use 
    138    an invoke with a normal destination which contains an unreachable 
    139    instruction.  This form conveys to the optimizer that the call returns 
    140    abnormally.  For an invoke which neither returns normally or requires unwind
    141    code in the current function, you can use a noreturn call instruction if 
    142    desired.  This is generally not required because the optimizer will convert
    143    an invoke with an unreachable unwind destination to a call instruction.
    144 
    145 #. Use profile metadata to indicate statically known cold paths, even if 
    146    dynamic profiling information is not available.  This can make a large 
    147    difference in code placement and thus the performance of tight loops.
    148 
    149 #. When generating code for loops, try to avoid terminating the header block of
    150    the loop earlier than necessary.  If the terminator of the loop header 
    151    block is a loop exiting conditional branch, the effectiveness of LICM will
    152    be limited for loads not in the header.  (This is due to the fact that LLVM 
    153    may not know such a load is safe to speculatively execute and thus can't 
    154    lift an otherwise loop invariant load unless it can prove the exiting 
    155    condition is not taken.)  It can be profitable, in some cases, to emit such 
    156    instructions into the header even if they are not used along a rarely 
    157    executed path that exits the loop.  This guidance specifically does not 
    158    apply if the condition which terminates the loop header is itself invariant,
    159    or can be easily discharged by inspecting the loop index variables.
    160 
    161 #. In hot loops, consider duplicating instructions from small basic blocks 
    162    which end in highly predictable terminators into their successor blocks.  
    163    If a hot successor block contains instructions which can be vectorized 
    164    with the duplicated ones, this can provide a noticeable throughput
    165    improvement.  Note that this is not always profitable and does involve a 
    166    potentially large increase in code size.
    167 
    168 #. When checking a value against a constant, emit the check using a consistent
    169    comparison type.  The GVN pass *will* optimize redundant equalities even if
    170    the type of comparison is inverted, but GVN only runs late in the pipeline.
    171    As a result, you may miss the opportunity to run other important 
    172    optimizations.  Improvements to EarlyCSE to remove this issue are tracked in 
    173    Bug 23333.
    174 
    175 #. Avoid using arithmetic intrinsics unless you are *required* by your source 
    176    language specification to emit a particular code sequence.  The optimizer 
    177    is quite good at reasoning about general control flow and arithmetic, it is
    178    not anywhere near as strong at reasoning about the various intrinsics.  If 
    179    profitable for code generation purposes, the optimizer will likely form the 
    180    intrinsics itself late in the optimization pipeline.  It is *very* rarely 
    181    profitable to emit these directly in the language frontend.  This item
    182    explicitly includes the use of the :ref:`overflow intrinsics <int_overflow>`.
    183 
    184 #. Avoid using the :ref:`assume intrinsic <int_assume>` until you've 
    185    established that a) there's no other way to express the given fact and b) 
    186    that fact is critical for optimization purposes.  Assumes are a great 
    187    prototyping mechanism, but they can have negative effects on both compile 
    188    time and optimization effectiveness.  The former is fixable with enough 
    189    effort, but the later is fairly fundamental to their designed purpose.
    190 
    191 
    192 Describing Language Specific Properties
    193 =======================================
    194 
    195 When translating a source language to LLVM, finding ways to express concepts 
    196 and guarantees available in your source language which are not natively 
    197 provided by LLVM IR will greatly improve LLVM's ability to optimize your code. 
    198 As an example, C/C++'s ability to mark every add as "no signed wrap (nsw)" goes
    199 a long way to assisting the optimizer in reasoning about loop induction 
    200 variables and thus generating more optimal code for loops.  
    201 
    202 The LLVM LangRef includes a number of mechanisms for annotating the IR with 
    203 additional semantic information.  It is *strongly* recommended that you become 
    204 highly familiar with this document.  The list below is intended to highlight a 
    205 couple of items of particular interest, but is by no means exhaustive.
    206 
    207 Restricted Operation Semantics
    208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    209 #. Add nsw/nuw flags as appropriate.  Reasoning about overflow is 
    210    generally hard for an optimizer so providing these facts from the frontend 
    211    can be very impactful.  
    212 
    213 #. Use fast-math flags on floating point operations if legal.  If you don't 
    214    need strict IEEE floating point semantics, there are a number of additional 
    215    optimizations that can be performed.  This can be highly impactful for 
    216    floating point intensive computations.
    217 
    218 Describing Aliasing Properties
    219 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    220 
    221 #. Add noalias/align/dereferenceable/nonnull to function arguments and return 
    222    values as appropriate
    223 
    224 #. Use pointer aliasing metadata, especially tbaa metadata, to communicate 
    225    otherwise-non-deducible pointer aliasing facts
    226 
    227 #. Use inbounds on geps.  This can help to disambiguate some aliasing queries.
    228 
    229 
    230 Modeling Memory Effects
    231 ^^^^^^^^^^^^^^^^^^^^^^^^
    232 
    233 #. Mark functions as readnone/readonly/argmemonly or noreturn/nounwind when
    234    known.  The optimizer will try to infer these flags, but may not always be
    235    able to.  Manual annotations are particularly important for external 
    236    functions that the optimizer can not analyze.
    237 
    238 #. Use the lifetime.start/lifetime.end and invariant.start/invariant.end 
    239    intrinsics where possible.  Common profitable uses are for stack like data 
    240    structures (thus allowing dead store elimination) and for describing 
    241    life times of allocas (thus allowing smaller stack sizes).  
    242 
    243 #. Mark invariant locations using !invariant.load and TBAA's constant flags
    244 
    245 Pass Ordering
    246 ^^^^^^^^^^^^^
    247 
    248 One of the most common mistakes made by new language frontend projects is to 
    249 use the existing -O2 or -O3 pass pipelines as is.  These pass pipelines make a
    250 good starting point for an optimizing compiler for any language, but they have 
    251 been carefully tuned for C and C++, not your target language.  You will almost 
    252 certainly need to use a custom pass order to achieve optimal performance.  A 
    253 couple specific suggestions:
    254 
    255 #. For languages with numerous rarely executed guard conditions (e.g. null 
    256    checks, type checks, range checks) consider adding an extra execution or 
    257    two of LoopUnswith and LICM to your pass order.  The standard pass order, 
    258    which is tuned for C and C++ applications, may not be sufficient to remove 
    259    all dischargeable checks from loops.
    260 
    261 #. If you language uses range checks, consider using the IRCE pass.  It is not 
    262    currently part of the standard pass order.
    263 
    264 #. A useful sanity check to run is to run your optimized IR back through the 
    265    -O2 pipeline again.  If you see noticeable improvement in the resulting IR, 
    266    you likely need to adjust your pass order.
    267 
    268 
    269 I Still Can't Find What I'm Looking For
    270 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    271 
    272 If you didn't find what you were looking for above, consider proposing an piece
    273 of metadata which provides the optimization hint you need.  Such extensions are
    274 relatively common and are generally well received by the community.  You will 
    275 need to ensure that your proposal is sufficiently general so that it benefits 
    276 others if you wish to contribute it upstream.
    277 
    278 You should also consider describing the problem you're facing on `llvm-dev 
    279 <http://lists.llvm.org/mailman/listinfo/llvm-dev>`_ and asking for advice.  
    280 It's entirely possible someone has encountered your problem before and can 
    281 give good advice.  If there are multiple interested parties, that also 
    282 increases the chances that a metadata extension would be well received by the
    283 community as a whole.  
    284 
    285 Adding to this document
    286 =======================
    287 
    288 If you run across a case that you feel deserves to be covered here, please send
    289 a patch to `llvm-commits
    290 <http://lists.llvm.org/mailman/listinfo/llvm-commits>`_ for review.
    291 
    292 If you have questions on these items, please direct them to `llvm-dev 
    293 <http://lists.llvm.org/mailman/listinfo/llvm-dev>`_.  The more relevant 
    294 context you are able to give to your question, the more likely it is to be 
    295 answered.
    296 
    297