Remarks

Introduction to the LLVM remark diagnostics

LLVM is able to emit diagnostics from passes describing whether an optimizationhas been performed or missed for a particular reason, which should give moreinsight to users about what the compiler did during the compilation pipeline.

There are three main remark types:

Passed

Remarks that describe a successful optimization performed by the compiler.

Example:
  1. foo inlined into bar with (cost=always): always inline attribute

Missed

Remarks that describe an attempt to an optimization by the compiler thatcould not be performed.

Example:
  1. foo not inlined into bar because it should never be inlined(cost=never): noinline function attribute

Analysis

Remarks that describe the result of an analysis, that can bring moreinformation to the user regarding the generated code.

Example:
  1. 16 stack bytes in function
  1. 10 instructions in function

Enabling optimization remarks

There are two modes that are supported for enabling optimization remarks inLLVM: through remark diagnostics, or through serialized remarks.

Remark diagnostics

Optimization remarks can be emitted as diagnostics. These diagnostics will bepropagated to front-ends if desired, or emitted by tools like llc or opt.

  • -pass-remarks=<regex>
  • Enables optimization remarks from passes whose name match the given (POSIX)regular expression.
  • -pass-remarks-missed=<regex>
  • Enables missed optimization remarks from passes whose name match the given(POSIX) regular expression.
  • -pass-remarks-analysis=<regex>
  • Enables optimization analysis remarks from passes whose name match the given(POSIX) regular expression.

Serialized remarks

While diagnostics are useful during development, it is often more useful torefer to optimization remarks post-compilation, typically during performanceanalysis.

For that, LLVM can serialize the remarks produced for each compilation unit toa file that can be consumed later.

By default, the format of the serialized remarks is YAML, and it can be accompanied by a sectionin the object files to easily retrieve it.

llc and opt support thefollowing options:

Basic options

-pass-remarks-output=<filename>

Enables the serialization of remarks to a file specified in <filename>.

By default, the output is serialized to YAML.

-pass-remarks-format=<format>

Specifies the output format of the serialized remarks.

Supported formats:

Content configuration

-pass-remarks-filter=<regex>

Only passes whose name match the given (POSIX) regular expression will beserialized to the final output.

-pass-remarks-with-hotness

With PGO, include profile count in optimization remarks.

-pass-remarks-hotness-threshold

The minimum profile count required for an optimization remark to beemitted.

Other tools that support remarks:

llvm-lto

-lto-pass-remarks-output=<filename>
-lto-pass-remarks-filter=<regex>
-lto-pass-remarks-format=<format>
-lto-pass-remarks-with-hotness
-lto-pass-remarks-hotness-threshold

gold-plugin and lld

-opt-remarks-filename=<filename>
-opt-remarks-filter=<regex>
-opt-remarks-format=<format>
-opt-remarks-with-hotness

Serialization modes

There are two modes available for serializing remarks:

Separate

In this mode, the remarks and the metadata are serialized separately. Theclient is responsible for parsing the metadata first, then use the metadatato correctly parse the remarks.

Standalone

In this mode, the remarks and the metadata are serialized to the samestream. The metadata will always come before the remarks.

The compiler does not support emitting standalone remarks. This mode ismore suited for post-processing tools like linkers, that can merge theremarks for one whole project.

YAML remarks

A typical remark serialized to YAML looks like this:

  1. --- !<TYPE>
  2. Pass: <pass>
  3. Name: <name>
  4. DebugLoc: { File: <file>, Line: <line>, Column: <column> }
  5. Function: <function>
  6. Hotness: <hotness>
  7. Args:
  8.   - <key>: <value>
  9.     DebugLoc: { File: <arg-file>, Line: <arg-line>, Column: <arg-column> }

The following entries are mandatory:

  • <TYPE>: can be Passed, Missed, Analysis,AnalysisFPCommute, AnalysisAliasing, Failure.
  • <pass>: the name of the pass that emitted this remark.
  • <name>: the name of the remark coming from <pass>.
  • <function>: the mangled name of the function.

If a DebugLoc entry is specified, the following fields are required:

  • <file>
  • <line>
  • <column>

If an arg entry is specified, the following fields are required:

  • <key>
  • <value>

If a DebugLoc entry is specified within an arg entry, the followingfields are required:

  • <arg-file>
  • <arg-line>
  • <arg-column>

YAML with a string table

The YAML serialization supports the usage of a string table by using theyaml-strtab format.

This format replaces strings in the YAML output with integers representing theindex in the string table that can be provided separately through metadata.

The following entries can take advantage of the string table while respectingYAML rules:

  • <pass>
  • <name>
  • <function>
  • <file>
  • <value>
  • <arg-file>

Currently, none of the tools in the opt-viewer directorysupport this format.

YAML metadata

The metadata used together with the YAML format is:

  • a magic number: “REMARKS\0”
  • the version number: a little-endian uint64_t
  • the total size of the string table (the size itself excluded):little-endian uint64_t
  • a list of null-terminated strings

Optional:

  • the absolute file path to the serialized remark diagnostics: anull-terminated string.

When the metadata is serialized separately from the remarks, the file pathshould be present and point to the file where the remarks are serialized to.

In case the metadata only acts as a header to the remarks, the file path can beomitted.

LLVM bitstream remarks

This format is using LLVM bitstream to serialize remarksand their associated metadata.

A bitstream remark stream can be identified by the magic number "RMRK" thatis placed at the very beginning.

The format for serializing remarks is composed of two different block types:

META_BLOCK

The block providing information about the rest of the content in the stream.

Exactly one block is expected. Having multiple metadata blocks is an error.

This block can contain the following records:

RECORD_META_CONTAINER_INFO

The container version and type.

Version: u32

Type: u2

RECORD_META_REMARK_VERSION

The version of the remark entries. This can change independently from thecontainer version.

Version: u32

RECORD_META_STRTAB

The string table used by the remark entries. The format of the string tableis a sequence of strings separated by \0.

RECORD_META_EXTERNAL_FILE

The external remark file path that contains the remark blocks associatedwith this metadata. This is an absolute path.

REMARK_BLOCK

The block describing a remark entry.

0 or more blocks per file are allowed. Each block will depend on theMETA_BLOCK in order to be parsed correctly.

This block can contain the following records:

RECORD_REMARK_HEADER

The header of the remark. This contains all the mandatory information abouta remark.

Typeu3
Remark nameVBR6 (string table index)
Pass nameVBR6 (string table index)
Function nameVBR6 (string table index)

RECORD_REMARK_DEBUG_LOC

The source location for the corresponding remark. This record is optional.

FileVBR7 (string table index)
Lineu32
Columnu32

RECORD_REMARK_HOTNESS

The hotness of the remark. This record is optional.

Hotness | VBR8 (string table index)

RECORD_REMARK_ARG_WITH_DEBUGLOC

A remark argument with an associated debug location.

KeyVBR7 (string table index)
ValueVBR7 (string table index)
FileVBR7 (string table index)
Lineu32
Columnu32

RECORD_REMARK_ARG_WITHOUT_DEBUGLOC

A remark argument with an associated debug location.

KeyVBR7 (string table index)
ValueVBR7 (string table index)

The remark container

Bitstream remarks are designed to be used in two different modes:

The separate mode

The separate mode is the mode that is typically used during compilation. Itprovides a way to serialize the remark entries to a stream while somemetadata is kept in memory to be emitted in the product of the compilation(typically, an object file).

The standalone mode

The standalone mode is typically stored and used after the distribution ofa program. It contains all the information that allows the parsing of allthe remarks without having any external dependencies.

In order to support multiple modes, the format introduces the concept of abitstream remark container type.

SeparateRemarksMeta: the metadata emitted separately

This container type expects only a META_BLOCK containing only:

Typically, this is emitted in a section in the object files, allowingclients to retrieve remarks and their associated metadata directly fromintermediate products.

SeparateRemarksFile: the remark entries emitted separately

This container type expects only a META_BLOCK containing only:

This container type expects 0 or more REMARK_BLOCK.

Typically, this is emitted in a side-file alongside an object file, and ismade to be able to stream to without increasing the memory consumption ofthe compiler. This is referenced by the RECORD_META_EXTERNAL_FILE entry in theSeparateRemarksMeta container.

When the parser tries to parse a container that contains the metadata for theseparate remarks, it should parse the version and type, then keep the stringtable in memory while opening the external file, validating its metadata andparsing the remark entries.

The container versions from the separate container should match in order tohave a well-formed file.

Standalone: the metadata and the remark entries emitted together

This container type expects only a META_BLOCK containing only:

This container type expects 0 or more REMARK_BLOCK.

A complete output of llvm-bcanalyzer on the different container types:

SeparateRemarksMeta

  1. <BLOCKINFO_BLOCK/>
  2. <Meta BlockID=8 NumWords=13 BlockCodeSize=3>
  3.   <Container info codeid=1 abbrevid=4 op0=5 op1=0/>
  4.   <String table codeid=3 abbrevid=5/> blob data = 'pass\\x00key\\x00value\\x00'
  5.   <External File codeid=4 abbrevid=6/> blob data = '/path/to/file/name'
  6. </Meta>

SeparateRemarksFile

  1. <BLOCKINFO_BLOCK/>
  2. <Meta BlockID=8 NumWords=3 BlockCodeSize=3>
  3.   <Container info codeid=1 abbrevid=4 op0=0 op1=1/>
  4.   <Remark version codeid=2 abbrevid=5 op0=0/>
  5. </Meta>
  6. <Remark BlockID=9 NumWords=8 BlockCodeSize=4>
  7.   <Remark header codeid=5 abbrevid=4 op0=2 op1=0 op2=1 op3=2/>
  8.   <Remark debug location codeid=6 abbrevid=5 op0=3 op1=99 op2=55/>
  9.   <Remark hotness codeid=7 abbrevid=6 op0=999999999/>
  10.   <Argument with debug location codeid=8 abbrevid=7 op0=4 op1=5 op2=6 op3=11 op4=66/>
  11. </Remark>

Standalone

  1. <BLOCKINFO_BLOCK/>
  2. <Meta BlockID=8 NumWords=15 BlockCodeSize=3>
  3.   <Container info codeid=1 abbrevid=4 op0=5 op1=2/>
  4.   <Remark version codeid=2 abbrevid=5 op0=30/>
  5.   <String table codeid=3 abbrevid=6/> blob data = 'pass\\x00remark\\x00function\\x00path\\x00key\\x00value\\x00argpath\\x00'
  6. </Meta>
  7. <Remark BlockID=9 NumWords=8 BlockCodeSize=4>
  8.   <Remark header codeid=5 abbrevid=4 op0=2 op1=1 op2=0 op3=2/>
  9.   <Remark debug location codeid=6 abbrevid=5 op0=3 op1=99 op2=55/>
  10.   <Remark hotness codeid=7 abbrevid=6 op0=999999999/>
  11.   <Argument with debug location codeid=8 abbrevid=7 op0=4 op1=5 op2=6 op3=11 op4=66/>
  12. </Remark>

opt-viewer

The opt-viewer directory contains a collection of tools that visualize andsummarize serialized remarks.

The tools only support the yaml format.

opt-viewer.py

Output a HTML page which gives visual feedback on compiler interactions withyour program.

Examples:
  1. $ opt-viewer.py my_yaml_file.opt.yaml
  1. $ opt-viewer.py my_build_dir/

opt-stats.py

Output statistics about the optimization remarks in the input set.

Example:
  1. $ opt-stats.py my_yaml_file.opt.yamlTotal number of remarks           3Top 10 remarks by pass:  inline                         33%  asm-printer                    33%  prologepilog                   33%Top 10 remarks:  asm-printer/InstructionCount   33%  inline/NoDefinition            33%  prologepilog/StackSize         33%

opt-diff.py

Produce a new YAML file which contains all of the changes in optimizationsbetween two YAML files.

Typically, this tool should be used to do diffs between:

  • new compiler + fixed source vs old compiler + fixed source
  • fixed compiler + new source vs fixed compiler + old source

This diff file can be displayed using opt-viewer.py.

Example:
  1. $ opt-diff.py my_opt_yaml1.opt.yaml my_opt_yaml2.opt.yaml -o my_opt_diff.opt.yaml$ opt-viewer.py my_opt_diff.opt.yaml

Emitting remark diagnostics in the object file

A section containing metadata on remark diagnostics will be emitted for thefollowing formats:

  • yaml-strtab
  • bitstream

This can be overridden by using the flag -remarks-section=<bool>.

The section is named:

  • LLVM,remarks (MachO)

C API

LLVM provides a library that can be used to parse remarks through a sharedlibrary named libRemarks.

The typical usage through the C API is like the following:

  1. LLVMRemarkParserRef Parser = LLVMRemarkParserCreateYAML(Buf, Size);
  2. LLVMRemarkEntryRef Remark = NULL;
  3. while ((Remark = LLVMRemarkParserGetNext(Parser))) {
  4.    // use Remark
  5.    LLVMRemarkEntryDispose(Remark); // Release memory.
  6. }
  7. bool HasError = LLVMRemarkParserHasError(Parser);
  8. LLVMRemarkParserDispose(Parser);

Remark streamers

The RemarkStreamer interface is used to unify the serializationcapabilities of remarks across all the components that can generate remarks.

All remark serialization should go through the main remark streamer, thellvm::remarks::RemarkStreamer set up in the LLVMContext. The interfacetakes remark objects converted to llvm::remarks::Remark, and takes care ofserializing it to the requested format, using the requested type of metadata,etc.

Typically, a specialized remark streamer will hold a reference to the one setup in the LLVMContext, and will operate on its own type of diagnostics.

For example, LLVM IR passes will emit llvm::DiagnosticInfoOptimization*that get converted to llvm::remarks::Remark objects. Then, clang could setup its own specialized remark streamer that takes clang::Diagnosticobjects. This can allow various components of the frontend to emit remarksusing the same techniques as the LLVM remarks.

This gives us the following advantages:

  • Composition: during the compilation pipeline, multiple components can set uptheir specialized remark streamers that all emit remarks through the samemain streamer.
  • Re-using the remark infrastructure in lib/Remarks.
  • Using the same file and format for the remark emitters created throughout thecompilation.

at the cost of an extra layer of abstraction.