Performance Tips for Frontend Authors

Abstract

The intended audience of this document is developers of language frontends targeting LLVM IR. This document is home to a collection of tips on how to generate IR that optimizes well.

IR Best Practices

As with any optimizer, LLVM has its strengths and weaknesses. In some cases, surprisingly small changes in the source IR can have a large effect on the generated code.

Beyond the specific items on the list below, it’s worth noting that the most 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 can often be useful to write a quick C program with the semantics you’re trying to model and see what decisions Clang’s IRGen makes about what IR to emit. Studying Clang’s CodeGen directory can also be a good source of ideas. Note that Clang and LLVM are explicitly version locked so you’ll need to make sure you’re using a Clang built from the same svn revision or release as the LLVM library you’re using. As always, it’s strongly recommended that you track tip of tree development, particularly during bring up of a new project.

The Basics

  1. Make sure that your Modules contain both a data layout specification and target triple. Without these pieces, non of the target specific optimization will be enabled. This can have a major effect on the generated code quality.
  2. For each function or global emitted, use the most private linkage type possible (private, internal or linkonce_odr preferably). Doing so will make LLVM’s inter-procedural optimizations much more effective.
  3. Avoid high in-degree basic blocks (e.g. basic blocks with dozens or hundreds of predecessors). Among other issues, the register allocator is known to perform badly with confronted with such structures. The only exception to this guidance is that a unified return block with high in-degree is fine.

Use of allocas

An alloca instruction can be used to represent a function scoped stack slot, but can also represent dynamic frame expansion. When representing function scoped variables or locations, placing alloca instructions at the beginning of the entry block should be preferred. In particular, place them before any call instructions. Call instructions might get inlined and replaced with multiple basic blocks. The end result is that a following alloca instruction would no longer be in the entry basic block afterward.

The SROA (Scalar Replacement Of Aggregates) and Mem2Reg passes only attempt to eliminate alloca instructions that are in the entry basic block. Given SSA is the canonical form expected by much of the optimizer; if allocas can not be eliminated by Mem2Reg or SROA, the optimizer is likely to be less effective than it could be.

Avoid loads and stores of large aggregate type

LLVM currently does not optimize well loads and stores of large aggregate types (i.e. structs and arrays). As an alternative, consider loading individual fields from memory.

Aggregates that are smaller than the largest (performant) load or store instruction supported by the targeted hardware are well supported. These can be an effective way to represent collections of small packed fields.

Zext GEP indices to machine register width

Internally, LLVM often promotes the width of GEP indices to machine register width. When it does so, it will default to using sign extension (sext) operations for safety. If your source language provides information about the range of the index, you may wish to manually extend indices to machine register width using a zext instruction.

When to specify alignment

LLVM will always generate correct code if you don’t specify alignment, but may generate inefficient code. For example, if you are targeting MIPS (or older ARM ISAs) then the hardware does not handle unaligned loads and stores, and so you will enter a trap-and-emulate path if you do a load or store with lower-than-natural alignment. To avoid this, LLVM will emit a slower sequence of loads, shifts and masks (or load-right + load-left on MIPS) for all cases where the load / store does not have a sufficiently high alignment in the IR.

The alignment is used to guarantee the alignment on allocas and globals, though in most cases this is unnecessary (most targets have a sufficiently high default alignment that they’ll be fine). It is also used to provide a contract to the back end saying ‘either this load/store has this alignment, or it is undefined behavior’. This means that the back end is free to emit instructions that rely on that alignment (and mid-level optimizers are free to perform transforms that require that alignment). For x86, it doesn’t make much difference, as almost all instructions are alignment-independent. For MIPS, it can make a big difference.

Note that if your loads and stores are atomic, the backend will be unable to lower an under aligned access into a sequence of natively aligned accesses. As a result, alignment is mandatory for atomic loads and stores.

Other Things to Consider

  1. Use ptrtoint/inttoptr sparingly (they interfere with pointer aliasing analysis), prefer GEPs
  2. Prefer globals over inttoptr of a constant address - this gives you dereferencability information. In MCJIT, use getSymbolAddress to provide actual address.
  3. Be wary of ordered and atomic memory operations. They are hard to optimize and may not be well optimized by the current optimizer. Depending on your source language, you may consider using fences instead.
  4. If calling a function which is known to throw an exception (unwind), use an invoke with a normal destination which contains an unreachable instruction. This form conveys to the optimizer that the call returns abnormally. For an invoke which neither returns normally or requires unwind code in the current function, you can use a noreturn call instruction if desired. This is generally not required because the optimizer will convert an invoke with an unreachable unwind destination to a call instruction.
  5. Use profile metadata to indicate statically known cold paths, even if dynamic profiling information is not available. This can make a large difference in code placement and thus the performance of tight loops.
  6. When generating code for loops, try to avoid terminating the header block of the loop earlier than necessary. If the terminator of the loop header block is a loop exiting conditional branch, the effectiveness of LICM will be limited for loads not in the header. (This is due to the fact that LLVM may not know such a load is safe to speculatively execute and thus can’t lift an otherwise loop invariant load unless it can prove the exiting condition is not taken.) It can be profitable, in some cases, to emit such instructions into the header even if they are not used along a rarely executed path that exits the loop. This guidance specifically does not apply if the condition which terminates the loop header is itself invariant, or can be easily discharged by inspecting the loop index variables.
  7. In hot loops, consider duplicating instructions from small basic blocks which end in highly predictable terminators into their successor blocks. If a hot successor block contains instructions which can be vectorized with the duplicated ones, this can provide a noticeable throughput improvement. Note that this is not always profitable and does involve a potentially large increase in code size.
  8. When checking a value against a constant, emit the check using a consistent comparison type. The GVN pass will optimize redundant equalities even if the type of comparison is inverted, but GVN only runs late in the pipeline. As a result, you may miss the opportunity to run other important optimizations. Improvements to EarlyCSE to remove this issue are tracked in Bug 23333.
  9. Avoid using arithmetic intrinsics unless you are required by your source language specification to emit a particular code sequence. The optimizer is quite good at reasoning about general control flow and arithmetic, it is not anywhere near as strong at reasoning about the various intrinsics. If profitable for code generation purposes, the optimizer will likely form the intrinsics itself late in the optimization pipeline. It is very rarely profitable to emit these directly in the language frontend. This item explicitly includes the use of the overflow intrinsics.
  10. Avoid using the assume intrinsic until you’ve established that a) there’s no other way to express the given fact and b) that fact is critical for optimization purposes. Assumes are a great prototyping mechanism, but they can have negative effects on both compile time and optimization effectiveness. The former is fixable with enough effort, but the later is fairly fundamental to their designed purpose.

Describing Language Specific Properties

When translating a source language to LLVM, finding ways to express concepts and guarantees available in your source language which are not natively provided by LLVM IR will greatly improve LLVM’s ability to optimize your code. As an example, C/C++’s ability to mark every add as “no signed wrap (nsw)” goes a long way to assisting the optimizer in reasoning about loop induction variables and thus generating more optimal code for loops.

The LLVM LangRef includes a number of mechanisms for annotating the IR with additional semantic information. It is strongly recommended that you become highly familiar with this document. The list below is intended to highlight a couple of items of particular interest, but is by no means exhaustive.

Restricted Operation Semantics

  1. Add nsw/nuw flags as appropriate. Reasoning about overflow is generally hard for an optimizer so providing these facts from the frontend can be very impactful.
  2. Use fast-math flags on floating point operations if legal. If you don’t need strict IEEE floating point semantics, there are a number of additional optimizations that can be performed. This can be highly impactful for floating point intensive computations.

Describing Aliasing Properties

  1. Add noalias/align/dereferenceable/nonnull to function arguments and return values as appropriate
  2. Use pointer aliasing metadata, especially tbaa metadata, to communicate otherwise-non-deducible pointer aliasing facts
  3. Use inbounds on geps. This can help to disambiguate some aliasing queries.

Modeling Memory Effects

  1. Mark functions as readnone/readonly/argmemonly or noreturn/nounwind when known. The optimizer will try to infer these flags, but may not always be able to. Manual annotations are particularly important for external functions that the optimizer can not analyze.
  2. Use the lifetime.start/lifetime.end and invariant.start/invariant.end intrinsics where possible. Common profitable uses are for stack like data structures (thus allowing dead store elimination) and for describing life times of allocas (thus allowing smaller stack sizes).
  3. Mark invariant locations using !invariant.load and TBAA’s constant flags

Pass Ordering

One of the most common mistakes made by new language frontend projects is to use the existing -O2 or -O3 pass pipelines as is. These pass pipelines make a good starting point for an optimizing compiler for any language, but they have been carefully tuned for C and C++, not your target language. You will almost certainly need to use a custom pass order to achieve optimal performance. A couple specific suggestions:

  1. For languages with numerous rarely executed guard conditions (e.g. null checks, type checks, range checks) consider adding an extra execution or two of LoopUnswith and LICM to your pass order. The standard pass order, which is tuned for C and C++ applications, may not be sufficient to remove all dischargeable checks from loops.
  2. If you language uses range checks, consider using the IRCE pass. It is not currently part of the standard pass order.
  3. A useful sanity check to run is to run your optimized IR back through the -O2 pipeline again. If you see noticeable improvement in the resulting IR, you likely need to adjust your pass order.

I Still Can’t Find What I’m Looking For

If you didn’t find what you were looking for above, consider proposing an piece of metadata which provides the optimization hint you need. Such extensions are relatively common and are generally well received by the community. You will need to ensure that your proposal is sufficiently general so that it benefits others if you wish to contribute it upstream.

You should also consider describing the problem you’re facing on llvm-dev and asking for advice. It’s entirely possible someone has encountered your problem before and can give good advice. If there are multiple interested parties, that also increases the chances that a metadata extension would be well received by the community as a whole.

Adding to this document

If you run across a case that you feel deserves to be covered here, please send a patch to llvm-commits for review.

If you have questions on these items, please direct them to llvm-dev. The more relevant context you are able to give to your question, the more likely it is to be answered.