Add IR support, ELF section and user documentation for partitioning feature.

[pcc]

Summary

The partitioning feature was proposed here:
http://lists.llvm.org/pipermail/llvm-dev/2019-February/130583.html

This is mostly just documentation. The feature itself will be contributed
in subsequent patches.

[ruiu]

lld/docs/Partitions.rst

16

share a virtual address space at link time?

28–32

Each partition's entry point address is fixed at link-time, so I wonder why you chose to get the address of the entry point by name. If you use the entry point by address, you can possibly remove the symbol table, no?

[peter.smith]

I've put some suggestions for the documentation. Feel free to ignore them if you prefer the original. I can't see anything wrong with the LLVM changes at a cursory look.

lld/docs/Partitions.rst

11

Program is an interesting choice of word here. From the initial RFC I think this is largely aimed at executables. I don't think that it would be impossible to make this work with shared-libraries, however I'd expect the constraints would make it difficult to use effectively. I'm thinking of an executable calling a function in the shared library that happens to be in a partition not loaded.

Maybe replace program with executable if that is the only option.

17

I tend to think of an address as an absolute address, I think "fixed load address relative to the main partition" is meant to be an offset". A suggestion:
"and each loadable partition is assigned a fixed offset from the main partition."

20

I think that you could remove the "by address" if offset were used earlier.

26

Just a suggestion, to be used as the "entry points" for each partition.

66

If I've understood it correctly you have to know ahead of time which source files are in which partition, the linker won't automatically create these for you. May be worth putting a shell comment before the main.c and before the feature.c emphasising that these are likely to be two separate build steps. I'm thinking that a common tool flow would be to build each partition into a separate static archive.

68

Just a thought, is --gc-sections compulsory? I don't think that it is but if you are hanging this off the --gc-sections it might be easier to automatically enable it if LLD encounters a partition. Or maybe have a --combined-object option that forces --gc-sections, and if not present drops all the partition information and produces a regular output?

71

This is an llvm-objcopy thought, perhaps a --extract-partition=all option would be useful. In this case the name of the partition would be the name of the filename as in your example above.

84

Could offset be used instead of relptr?

[pcc]

• Address review comments

lld/docs/Partitions.rst

11

It can be used with both executables and shared libraries. Indeed, the native parts of Android apps are shared libraries, so the shared library case is important to me. I'll clarify this in the doc.

In your executable calling a shared library scenario, the executable is linked separately from the shared library, and depends on it in the usual way (either DT_NEEDED or dlopen), correct? In this case, I don't think it would be possible for such a call to happen. To support the intended usage model of looking up partition entry points by name, the dynamic symbol table of each partition (including the main partition) defines just the symbols for that partition.

One thing that could cause problems is if the executable had a DT_NEEDED dependency on a loadable partition, as this may result in the main partition being loaded at the wrong relative address. In this case, I think it should be possible for a dynamic loader to make this work by using something like the dynamic tag that I mentioned in my initial proposal.

16

Yes, that might be the best way to phrase it. The partitions share a virtual address space in the p_vaddr sense, but that could probably be confused with the runtime virtual address space which is shared by everything.

28–32

A couple of reasons:

• This forces users to explicitly load the partition in order to access the entry point. If the entry point addresses were available statically, it would be easier to accidentally refer to a partition's entry point before it is loaded.
• In some cases, it's desirable for the entry points to be findable by name in the dynamic symbol table. For example, this is how JNI finds implementations of native methods; see: https://docs.oracle.com/javase/7/docs/technotes/guides/jni/spec/design.html#wp615 which could be very useful for Android apps.

66

Done. Note that it might not work for them to be archives because that gives no guarantee that the archive members will be added to the link, especially not if there won't be direct references. (The same thing applies to anything else that might need to be exported from a shared object.) In Chromium we're going to use a source_set [1] for the partition entry points.

[1] https://gn.googlesource.com/gn/+/refs/heads/master/docs/reference.md#func_source_set

68

Just a thought, is --gc-sections compulsory?

It technically isn't, but the behaviour without it isn't all that useful. With my current implementation, what I think will happen is that the linker will put almost everything in the main partition, and the loadable partitions will just contain symbol tables (whose symbols will point into the main partition). So it might indeed be a good idea to enable --gc-sections if partitions are encountered.

Or maybe have a --combined-object option that forces --gc-sections, and if not present drops all the partition information and produces a regular output?

We might want an option for controlling whether to create a combined output file but I don't think we should have that to start with and I don't think we should be dropping the partition specifications by default. Similar to my point in http://lists.llvm.org/pipermail/llvm-dev/2019-March/131015.html , leaving them in would allow downstream tools to detect the case where partition specifications were present but not acted upon.

71

It might occasionally be useful (though I'd probably name it --extract-all-partitions to avoid the (admittedly corner case) ambiguity between extracting all partitions and the partition named "all"), but since in most cases, the call to llvm-objcopy will be integrated into a build system which will know exactly which partitions are being created (and should be able to schedule the llvm-objcopy commands in parallel), combined with the fact that this feature can be relatively easily implemented in a shell script, it doesn't seem that important to me.

84

Yes, that works too.

[peter.smith]

Thanks for the updates and clarifications, I don't have any more comments.

[bd1976llvm]

Hi Peter,

I really like the concept. I had some nebulous concerns when I read the RFC; but, I didn't comment as I couldn't think of any alternatives to what you were proposing. What I am writing here is still not fully formed.. so feel free to ignore if it is unhelpful.

My main concern is that this is supporting a rare usecase and I don't really like the idea of adding complexity to the core tools for niche cases.

I feel like it would be better if this could be implemented via post-processing in some way; outside of the core tools.

Some aspect of this, like the linker being able to report all of the functions reachable from a given set of entry points, are of general utility though.

I wonder if there is any scope for implementing something different along the lines of:

1. Do an initial link.
2. Analyse the resulting elf and report all of the functions in each partition.
3. Do another link but this time provide an ordering file so that all of the functions for each partition are placed contiguously.
4. Extract the functions for each partition, apart form the first, into a partition file which also contains metadata specifying the address range.
5. Write zeros over all of the functions apart from those in the main partition and then compress the executable (saving the file space).
6. A loader extension patches in the missing functions for each partition when needed.

[pcc]

In D60242#1456022, @bd1976llvm wrote:

Hi Peter,

I really like the concept. I had some nebulous concerns when I read the RFC; but, I didn't comment as I couldn't think of any alternatives to what you were proposing. What I am writing here is still not fully formed.. so feel free to ignore if it is unhelpful.

My main concern is that this is supporting a rare usecase and I don't really like the idea of adding complexity to the core tools for niche cases.

I feel like it would be better if this could be implemented via post-processing in some way; outside of the core tools.

Some aspect of this, like the linker being able to report all of the functions reachable from a given set of entry points, are of general utility though.

I wonder if there is any scope for implementing something different along the lines of:

1. Do an initial link.
2. Analyse the resulting elf and report all of the functions in each partition.
3. Do another link but this time provide an ordering file so that all of the functions for each partition are placed contiguously.
4. Extract the functions for each partition, apart form the first, into a partition file which also contains metadata specifying the address range.
5. Write zeros over all of the functions apart from those in the main partition and then compress the executable (saving the file space).
6. A loader extension patches in the missing functions for each partition when needed.

Hi Ben, thanks for the feedback. I shared a similar concern around complexity, but I don't see any reasonable alternatives to implementing this directly in the linker.

Your idea of using orderfiles is interesting, but I see at least two showstopper problems with it. The first is that it is incompatible with range extension thunks. Imagine that you have one main partition and two loadable partitions. An orderfile is used to sort the main partition's sections before partition 1's sections before partition 2's sections. We're going to split partition 1 and partition 2 into separate files, but the linker doesn't know that, so it may place a range extension thunk in the middle of partition 1's sections and have partition 2 call it. Now at runtime we will crash if partition 2 is loaded without partition 1 and it tries to call one of the range extension thunks in partition 1. We might be able to work around this issue somehow by telling the linker where the partition boundaries are, but at that point since the linker needs to know how you're going to split the file up anyway it might as well actually split it up for you.

The second is that replacing parts of the file with zeros and compressing it isn't always useful from a size perspective; in many cases it's the uncompressed size that matters because that's what takes up storage space on the device. One might consider uncompressing the code at program startup but that leads to other problems: it slows down startup and prevents the kernel from paging out the pages containing the uncompressed code.

I also see other problems that aren't showstoppers but are still significant:

• We're leaving some of the binary size gains on the table because we can't move exception-handling sections, dynamically relocatable sections (e.g. vtables) or symbol tables.
• Since we aren't splitting the symbol table the solution is more error prone since there's nothing stopping someone from dlsym'ing a symbol in an unloaded partition.
• This requires linking multiple times, which will slow down linking and hurt developer productivity, but the problem is made worse when (full) LTO is being used since we'll need to do LTO multiple times.

[pcc]

• Add missing pieces to MC; add test

[bd1976llvm]

In D60242#1456509, @pcc wrote:

In D60242#1456022, @bd1976llvm wrote:

Hi Peter,

I really like the concept. I had some nebulous concerns when I read the RFC; but, I didn't comment as I couldn't think of any alternatives to what you were proposing. What I am writing here is still not fully formed.. so feel free to ignore if it is unhelpful.

My main concern is that this is supporting a rare usecase and I don't really like the idea of adding complexity to the core tools for niche cases.

I feel like it would be better if this could be implemented via post-processing in some way; outside of the core tools.

Some aspect of this, like the linker being able to report all of the functions reachable from a given set of entry points, are of general utility though.

I wonder if there is any scope for implementing something different along the lines of:

1. Do an initial link.
2. Analyse the resulting elf and report all of the functions in each partition.
3. Do another link but this time provide an ordering file so that all of the functions for each partition are placed contiguously.
4. Extract the functions for each partition, apart form the first, into a partition file which also contains metadata specifying the address range.
5. Write zeros over all of the functions apart from those in the main partition and then compress the executable (saving the file space).
6. A loader extension patches in the missing functions for each partition when needed.

Hi Ben, thanks for the feedback. I shared a similar concern around complexity, but I don't see any reasonable alternatives to implementing this directly in the linker.

Your idea of using orderfiles is interesting, but I see at least two showstopper problems with it. The first is that it is incompatible with range extension thunks. Imagine that you have one main partition and two loadable partitions. An orderfile is used to sort the main partition's sections before partition 1's sections before partition 2's sections. We're going to split partition 1 and partition 2 into separate files, but the linker doesn't know that, so it may place a range extension thunk in the middle of partition 1's sections and have partition 2 call it. Now at runtime we will crash if partition 2 is loaded without partition 1 and it tries to call one of the range extension thunks in partition 1. We might be able to work around this issue somehow by telling the linker where the partition boundaries are, but at that point since the linker needs to know how you're going to split the file up anyway it might as well actually split it up for you.

The second is that replacing parts of the file with zeros and compressing it isn't always useful from a size perspective; in many cases it's the uncompressed size that matters because that's what takes up storage space on the device. One might consider uncompressing the code at program startup but that leads to other problems: it slows down startup and prevents the kernel from paging out the pages containing the uncompressed code.

I also see other problems that aren't showstoppers but are still significant:

• We're leaving some of the binary size gains on the table because we can't move exception-handling sections, dynamically relocatable sections (e.g. vtables) or symbol tables.
• Since we aren't splitting the symbol table the solution is more error prone since there's nothing stopping someone from dlsym'ing a symbol in an unloaded partition.
• This requires linking multiple times, which will slow down linking and hurt developer productivity, but the problem is made worse when (full) LTO is being used since we'll need to do LTO multiple times.

Thanks for considering this carefully. IMO most of your concerns could be accepted as reasonable limitations to keep the feature out of the linker. We could probably deal with the points that you raised with the compression scheme somehow (e.g. some horrible scheme that adjusts the existing program headers and adds an additional bss one, similar to your original proposal but done by post processing). However, I think your point about thunks is a hard showstopper. We would need a much more expressive means of controlling the linkers placement of things (than the current ordering files, linker scripts, --section-start etc..) in order to create a layout that would be amenable to post processing; and, as you hinted at, this would be just as much work (or more) than just getting the linker to do the splitting up.

Given the above, I am happy with the implementation.

[pcc]

Adding some more people who may be able to review the LLVM changes here.

[pcc]

Ping.

[pcc]

• Add a note about the limit on the number of partitions

[pcc]

Ping^2.

[pcc]

• Rebase

[echristo]

Some inline comment requests and if you wouldn't mind changing the type allocator separately it'd be great.

Otherwise LGTM.

Thanks!

llvm/include/llvm/IR/GlobalValue.h

111

Comment please.

llvm/lib/AsmParser/LLParser.cpp

959

I know nothing else does, but commenting this would be great with invariants, conditions, etc.

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

2979

Comment what the conditional is checking please?

3071

Comment what the conditional is checking please?

3151

Comment here too please.

[pcc]

Thanks for the review. Committed rL361922 (rename type allocator) and rL361923 (this change).

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

2979

I added some comments to this file, but I'm not sure how useful they are. Maybe I'm biased from working on this code before, but (aside from the additional comment I added in parseFunctionRecord) I think the code speaks for itself here and I'm not sure if my comment is really adding any useful information. If we added similar comments elsewhere in these functions for consistency, I think that would result in less information per line of code which can hurt readability. Can you see a better comment to add here?

lld/docs/Partitions.rst

+Partitions
+==========
+
+.. warning::
+
+  This feature has not yet fully landed in LLD. This document describes how
+  the feature is intended to work once it lands. Furthermore, the feature
+  is currently experimental, and its interface is subject to change.
+
+LLD's partitioning feature allows a program (which may be an executable
+or a shared library) to be split into multiple pieces, or partitions. A
+partitioned program consists of a main partition together with a number of
+loadable partitions. The loadable partitions depend on the main partition
+in a similar way to a regular ELF shared object dependency, but unlike a
+shared object, the main partition and the loadable partitions share a virtual
+address space at link time, and each loadable partition is assigned a fixed
+offset from the main partition. This allows the loadable partitions to refer
+to code and data in the main partition directly without the binary size and
+performance overhead of PLTs, GOTs or symbol table entries.
+
+Usage
+-----
+
+A program that uses the partitioning feature must decide which symbols are
+going to be used as the "entry points" for each partition. An entry point
+could, for example, be the equivalent of the partition's ``main`` function, or
+there could be a group of functions that expose the functionality implemented
+by the partition. The intent is that in order to use a loadable partition,
+the program will use ``dlopen``/``dlsym`` or similar functions to dynamically
+load the partition at its assigned address, look up an entry point by name
+and call it. Note, however, that the standard ``dlopen`` function does not
+allow specifying a load address. On Android, the ``android_dlopen_ext``
+function may be used together with the ``ANDROID_DLEXT_RESERVED_ADDRESS``
+flag to load a shared object at a specific address.
+
+Once the entry points have been decided, the translation unit(s)
+containing the entry points should be compiled using the Clang compiler flag
+``-fsymbol-partition=<soname>``, where ``<soname>`` is the intended soname
+of the partition. The resulting object files are passed to the linker in
+the usual way.
+
+The linker will then use these entry points to automatically split the program
+into partitions according to which sections of the program are reachable from
+which entry points, similarly to how ``--gc-sections`` removes unused parts of
+a program. Any sections that are only reachable from a loadable partition's
+entry point are assigned to that partition, while all other sections are
+assigned to the main partition, including sections only reachable from
+loadable partitions.
+
+The following diagram illustrates how sections are assigned to partitions. Each
+section is colored according to its assigned partition.
+
+.. image:: partitions.svg
+
+The result of linking a program that uses partitions is essentially an
+ELF file with all of the partitions concatenated together. This file is
+referred to as a combined output file. To extract a partition from the
+combined output file, the ``llvm-objcopy`` tool should be used together
+with the flag ``--extract-main-partition`` to extract the main partition, or
+``-extract-partition=<soname>`` to extract one of the loadable partitions.
+An example command sequence is shown below:
+
+.. code-block:: shell
+
+  # Compile the main program.
+  clang -ffunction-sections -fdata-sections -c main.c
+
+  # Compile a feature to be placed in a loadable partition.
+  # Note that this is likely to be a separate build step to the main partition.
+  clang -ffunction-sections -fdata-sections -fsymbol-partition=libfeature.so -c feature.c
+
+  # Link the combined output file.
+  clang main.o feature.o -fuse-ld=lld -shared -o libcombined.so -Wl,-soname,libmain.so -Wl,--gc-sections
+
+  # Extract the partitions.
+  llvm-objcopy libcombined.so libmain.so --extract-main-partition
+  llvm-objcopy libcombined.so libfeature.so --extract-partition=libfeature.so
+
+In order to allow a program to discover the names of its loadable partitions
+and the locations of their reserved regions, the linker creates a partition
+index, which is an array of structs with the following definition:
+
+.. code-block:: c
+
+  struct partition_index_entry {
+    int32_t name_offset;
+    int32_t addr_offset;
+    uint32_t size;
+  };
+
+The ``name_offset`` field is a relative pointer to a null-terminated string
+containing the soname of the partition, the ``addr_offset`` field is a
+relative pointer to its load address and the ``size`` field contains the
+size of the region reserved for the partition. To derive an absolute pointer
+from the relative pointer fields in this data structure, the address of the
+field should be added to the value stored in the field.
+
+The program may discover the location of the partition index using the
+linker-defined symbols ``__part_index_begin`` and ``__part_index_end``.
+
+Restrictions
+------------
+
+This feature is currently only supported in the ELF linker.
+
+The partitioning feature may not currently be used together with the
+``SECTIONS`` or ``PHDRS`` linker script features, nor may it be used with the
+``--section-start``, ``-Ttext``, ``-Tdata`` or ``-Tbss`` flags. All of these
+features assume a single set of output sections and/or program headers, which
+makes their semantics ambiguous in the presence of more than one partition.
+
+The partitioning feature may not currently be used on the MIPS architecture
+because it is unclear whether the MIPS multi-GOT ABI is compatible with
+partitions.

lld/docs/index.rst

 .. toctree::
    :maxdepth: 1
 
    NewLLD
    AtomLLD
    WebAssembly
    windows_support
    missingkeyfunction
+   Partitions
    ReleaseNotes

lld/docs/partitions.dot

+digraph G {
+  part_main [label="Main partition",shape=plaintext];
+  part1 [label="Loadable partition 1",shape=plaintext];
+  part2 [label="Loadable partition 2",shape=plaintext];
+  main [style=filled,fillcolor=lightblue];
+  f1 [style=filled,fillcolor=lightsalmon];
+  f2 [style=filled,fillcolor=palegreen];
+  f3 [style=filled,fillcolor=lightblue];
+  f4 [style=filled,fillcolor=lightsalmon];
+  f5 [style=filled,fillcolor=lightblue];
+  f6 [style=filled,fillcolor=palegreen];
+  part_main -> main;
+  main -> f3;
+  part1 -> f1;
+  f1 -> f3;
+  f1 -> f4;
+  f1 -> f5;
+  part2 -> f2;
+  f2 -> f3;
+  f2 -> f5;
+  f2 -> f6;
+}

lld/docs/partitions.svg

+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN"
+ "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
+<!-- Generated by graphviz version 2.38.0 (20140413.2041)
+ -->
+<!-- Title: G Pages: 1 -->
+<svg width="393pt" height="188pt"
+ viewBox="0.00 0.00 393.00 188.00" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink">
+<g id="graph0" class="graph" transform="scale(1 1) rotate(0) translate(4 184)">
+<title>G</title>
+<polygon fill="white" stroke="none" points="-4,4 -4,-184 389,-184 389,4 -4,4"/>
+<!-- part_main -->
+<g id="node1" class="node"><title>part_main</title>
+<text text-anchor="middle" x="47.5" y="-158.3" font-family="Times,serif" font-size="14.00">Main partition</text>
+</g>
+<!-- main -->
+<g id="node4" class="node"><title>main</title>
+<ellipse fill="lightblue" stroke="black" cx="75.5" cy="-90" rx="28.6953" ry="18"/>
+<text text-anchor="middle" x="75.5" y="-86.3" font-family="Times,serif" font-size="14.00">main</text>
+</g>
+<!-- part_main&#45;&gt;main -->
+<g id="edge1" class="edge"><title>part_main&#45;&gt;main</title>
+<path fill="none" stroke="black" d="M54.4214,-143.697C57.6218,-135.696 61.492,-126.02 65.0381,-117.155"/>
+<polygon fill="black" stroke="black" points="68.3868,-118.207 68.8511,-107.622 61.8874,-115.607 68.3868,-118.207"/>
+</g>
+<!-- part1 -->
+<g id="node2" class="node"><title>part1</title>
+<text text-anchor="middle" x="176.5" y="-158.3" font-family="Times,serif" font-size="14.00">Loadable partition 1</text>
+</g>
+<!-- f1 -->
+<g id="node5" class="node"><title>f1</title>
+<ellipse fill="lightsalmon" stroke="black" cx="176.5" cy="-90" rx="27" ry="18"/>
+<text text-anchor="middle" x="176.5" y="-86.3" font-family="Times,serif" font-size="14.00">f1</text>
+</g>
+<!-- part1&#45;&gt;f1 -->
+<g id="edge3" class="edge"><title>part1&#45;&gt;f1</title>
+<path fill="none" stroke="black" d="M176.5,-143.697C176.5,-135.983 176.5,-126.712 176.5,-118.112"/>
+<polygon fill="black" stroke="black" points="180,-118.104 176.5,-108.104 173,-118.104 180,-118.104"/>
+</g>
+<!-- part2 -->
+<g id="node3" class="node"><title>part2</title>
+<text text-anchor="middle" x="321.5" y="-158.3" font-family="Times,serif" font-size="14.00">Loadable partition 2</text>
+</g>
+<!-- f2 -->
+<g id="node6" class="node"><title>f2</title>
+<ellipse fill="palegreen" stroke="black" cx="284.5" cy="-90" rx="27" ry="18"/>
+<text text-anchor="middle" x="284.5" y="-86.3" font-family="Times,serif" font-size="14.00">f2</text>
+</g>
+<!-- part2&#45;&gt;f2 -->
+<g id="edge7" class="edge"><title>part2&#45;&gt;f2</title>
+<path fill="none" stroke="black" d="M312.354,-143.697C307.97,-135.403 302.636,-125.311 297.813,-116.187"/>
+<polygon fill="black" stroke="black" points="300.801,-114.35 293.034,-107.145 294.612,-117.621 300.801,-114.35"/>
+</g>
+<!-- f3 -->
+<g id="node7" class="node"><title>f3</title>
+<ellipse fill="lightblue" stroke="black" cx="104.5" cy="-18" rx="27" ry="18"/>
+<text text-anchor="middle" x="104.5" y="-14.3" font-family="Times,serif" font-size="14.00">f3</text>
+</g>
+<!-- main&#45;&gt;f3 -->
+<g id="edge2" class="edge"><title>main&#45;&gt;f3</title>
+<path fill="none" stroke="black" d="M82.3726,-72.411C85.7675,-64.2164 89.9422,-54.1395 93.7473,-44.9548"/>
+<polygon fill="black" stroke="black" points="97.0828,-46.0481 97.6767,-35.4699 90.6158,-43.3689 97.0828,-46.0481"/>
+</g>
+<!-- f1&#45;&gt;f3 -->
+<g id="edge4" class="edge"><title>f1&#45;&gt;f3</title>
+<path fill="none" stroke="black" d="M161.93,-74.8345C151.75,-64.9376 137.976,-51.5462 126.469,-40.3591"/>
+<polygon fill="black" stroke="black" points="128.905,-37.8461 119.296,-33.3847 124.026,-42.865 128.905,-37.8461"/>
+</g>
+<!-- f4 -->
+<g id="node8" class="node"><title>f4</title>
+<ellipse fill="lightsalmon" stroke="black" cx="176.5" cy="-18" rx="27" ry="18"/>
+<text text-anchor="middle" x="176.5" y="-14.3" font-family="Times,serif" font-size="14.00">f4</text>
+</g>
+<!-- f1&#45;&gt;f4 -->
+<g id="edge5" class="edge"><title>f1&#45;&gt;f4</title>
+<path fill="none" stroke="black" d="M176.5,-71.6966C176.5,-63.9827 176.5,-54.7125 176.5,-46.1124"/>
+<polygon fill="black" stroke="black" points="180,-46.1043 176.5,-36.1043 173,-46.1044 180,-46.1043"/>
+</g>
+<!-- f5 -->
+<g id="node9" class="node"><title>f5</title>
+<ellipse fill="lightblue" stroke="black" cx="248.5" cy="-18" rx="27" ry="18"/>
+<text text-anchor="middle" x="248.5" y="-14.3" font-family="Times,serif" font-size="14.00">f5</text>
+</g>
+<!-- f1&#45;&gt;f5 -->
+<g id="edge6" class="edge"><title>f1&#45;&gt;f5</title>
+<path fill="none" stroke="black" d="M191.07,-74.8345C201.25,-64.9376 215.024,-51.5462 226.531,-40.3591"/>
+<polygon fill="black" stroke="black" points="228.974,-42.865 233.704,-33.3847 224.095,-37.8461 228.974,-42.865"/>
+</g>
+<!-- f2&#45;&gt;f3 -->
+<g id="edge8" class="edge"><title>f2&#45;&gt;f3</title>
+<path fill="none" stroke="black" d="M260.806,-81.0022C232.063,-71.1346 182.266,-53.5073 140.5,-36 138.683,-35.2385 136.825,-34.4358 134.957,-33.6106"/>
+<polygon fill="black" stroke="black" points="136.231,-30.3452 125.68,-29.3829 133.328,-36.7149 136.231,-30.3452"/>
+</g>
+<!-- f2&#45;&gt;f5 -->
+<g id="edge9" class="edge"><title>f2&#45;&gt;f5</title>
+<path fill="none" stroke="black" d="M276.15,-72.7646C271.788,-64.2831 266.353,-53.7144 261.459,-44.1974"/>
+<polygon fill="black" stroke="black" points="264.49,-42.4395 256.804,-35.1473 258.265,-45.6409 264.49,-42.4395"/>
+</g>
+<!-- f6 -->
+<g id="node10" class="node"><title>f6</title>
+<ellipse fill="palegreen" stroke="black" cx="320.5" cy="-18" rx="27" ry="18"/>
+<text text-anchor="middle" x="320.5" y="-14.3" font-family="Times,serif" font-size="14.00">f6</text>
+</g>
+<!-- f2&#45;&gt;f6 -->
+<g id="edge10" class="edge"><title>f2&#45;&gt;f6</title>
+<path fill="none" stroke="black" d="M292.85,-72.7646C297.212,-64.2831 302.647,-53.7144 307.541,-44.1974"/>
+<polygon fill="black" stroke="black" points="310.735,-45.6409 312.196,-35.1473 304.51,-42.4395 310.735,-45.6409"/>
+</g>
+</g>
+</svg>

llvm/docs/Extensions.rst

 this directive, all symbols are considered address-significant.
 
 .. code-block:: gas
 
   .addrsig_sym sym
 
 This marks ``sym`` as address-significant.
 
+``SHT_LLVM_SYMPART`` Section (symbol partition specification)
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This section is used to mark symbols with the `partition`_ that they
+belong to. An ``.llvm_sympart`` section consists of a null-terminated string
+specifying the name of the partition followed by a relocation referring to
+the symbol that belongs to the partition. It may be constructed as follows:
+
+.. code-block:: gas
+
+  .section ".llvm_sympart","",@llvm_sympart
+  .asciz "libpartition.so"
+  .word symbol_in_partition
+
+.. _partition: https://lld.llvm.org/Partitions.html
+
 CodeView-Dependent
 ------------------
 
 ``.cv_file`` Directive
 ^^^^^^^^^^^^^^^^^^^^^^
 Syntax:
   ``.cv_file`` *FileNumber FileName* [ *checksum* ] [ *checksumkind* ]
 

llvm/include/llvm/BinaryFormat/ELF.h

enum : unsigned {
   // https://android.googlesource.com/platform/bionic/+/6f12bfece5dcc01325e0abba56a46b1bcf991c69/tools/relocation_packer/src/elf_file.cc#37
   SHT_ANDROID_REL = 0x60000001,
   SHT_ANDROID_RELA = 0x60000002,
   SHT_LLVM_ODRTAB = 0x6fff4c00,         // LLVM ODR table.
   SHT_LLVM_LINKER_OPTIONS = 0x6fff4c01, // LLVM Linker Options.
   SHT_LLVM_CALL_GRAPH_PROFILE = 0x6fff4c02, // LLVM Call Graph Profile.
   SHT_LLVM_ADDRSIG = 0x6fff4c03,        // List of address-significant symbols
                                         // for safe ICF.
+  SHT_LLVM_SYMPART = 0x6fff4c04,        // Symbol partition specification.
   // Android's experimental support for SHT_RELR sections.
   // https://android.googlesource.com/platform/bionic/+/b7feec74547f84559a1467aca02708ff61346d2a/libc/include/elf.h#512
   SHT_ANDROID_RELR = 0x6fffff00,        // Relocation entries; only offsets.
   SHT_GNU_ATTRIBUTES = 0x6ffffff5,      // Object attributes.
   SHT_GNU_HASH = 0x6ffffff6,            // GNU-style hash table.
   SHT_GNU_verdef = 0x6ffffffd,          // GNU version definitions.
   SHT_GNU_verneed = 0x6ffffffe,         // GNU version references.
   SHT_GNU_versym = 0x6fffffff,          // GNU symbol versions table.

llvm/include/llvm/IR/GlobalValue.h

 
 protected:
   GlobalValue(Type *Ty, ValueTy VTy, Use *Ops, unsigned NumOps,
               LinkageTypes Linkage, const Twine &Name, unsigned AddressSpace)
       : Constant(PointerType::get(Ty, AddressSpace), VTy, Ops, NumOps),
         ValueType(Ty), Visibility(DefaultVisibility),
         UnnamedAddrVal(unsigned(UnnamedAddr::None)),
         DllStorageClass(DefaultStorageClass), ThreadLocal(NotThreadLocal),
-        HasLLVMReservedName(false), IsDSOLocal(false), IntID((Intrinsic::ID)0U),
-        Parent(nullptr) {
+        HasLLVMReservedName(false), IsDSOLocal(false), HasPartition(false),
+        IntID((Intrinsic::ID)0U), Parent(nullptr) {
     setLinkage(Linkage);
     setName(Name);
   }
 
   Type *ValueType;
 
-  static const unsigned GlobalValueSubClassDataBits = 17;
+  static const unsigned GlobalValueSubClassDataBits = 16;
 
   // All bitfields use unsigned as the underlying type so that MSVC will pack
   // them.
   unsigned Linkage : 4;       // The linkage of this global
   unsigned Visibility : 2;    // The visibility style of this global
   unsigned UnnamedAddrVal : 2; // This value's address is not significant
   unsigned DllStorageClass : 2; // DLL storage class
 
   unsigned ThreadLocal : 3; // Is this symbol "Thread Local", if so, what is
                             // the desired model?
 
   /// True if the function's name starts with "llvm.".  This corresponds to the
   /// value of Function::isIntrinsic(), which may be true even if
   /// Function::intrinsicID() returns Intrinsic::not_intrinsic.
   unsigned HasLLVMReservedName : 1;
 
   /// If true then there is a definition within the same linkage unit and that
   /// definition cannot be runtime preempted.
   unsigned IsDSOLocal : 1;
 
+  unsigned HasPartition : 1;
+
 private:
   // Give subclasses access to what otherwise would be wasted padding.
-  // (17 + 4 + 2 + 2 + 2 + 3 + 1 + 1) == 32.
+  // (16 + 4 + 2 + 2 + 2 + 3 + 1 + 1 + 1) == 32.
   unsigned SubClassData : GlobalValueSubClassDataBits;
 
   friend class Constant;
 
   void destroyConstantImpl();
   Value *handleOperandChangeImpl(Value *From, Value *To);
 
   /// Returns true if the definition of this global may be replaced by a
public:
   Type *getValueType() const { return ValueType; }
 
   void setDSOLocal(bool Local) { IsDSOLocal = Local; }
 
   bool isDSOLocal() const {
     return IsDSOLocal;
   }
 
+  bool hasPartition() const {
+    return HasPartition;
+  }
+  StringRef getPartition() const;
+  void setPartition(StringRef Part);
+
   static LinkageTypes getLinkOnceLinkage(bool ODR) {
     return ODR ? LinkOnceODRLinkage : LinkOnceAnyLinkage;
   }
   static LinkageTypes getWeakLinkage(bool ODR) {
     return ODR ? WeakODRLinkage : WeakAnyLinkage;
   }
 
   static bool isExternalLinkage(LinkageTypes Linkage) {

llvm/lib/AsmParser/LLLexer.cpp

#define KEYWORD(STR)                                                           \
   KEYWORD(nuw);
   KEYWORD(nsw);
   KEYWORD(exact);
   KEYWORD(inbounds);
   KEYWORD(inrange);
   KEYWORD(align);
   KEYWORD(addrspace);
   KEYWORD(section);
+  KEYWORD(partition);
   KEYWORD(alias);
   KEYWORD(ifunc);
   KEYWORD(module);
   KEYWORD(asm);
   KEYWORD(sideeffect);
   KEYWORD(alignstack);
   KEYWORD(inteldialect);
   KEYWORD(gc);

llvm/lib/AsmParser/LLParser.cpp

GA.reset(GlobalIFunc::create(Ty, AddrSpace,
                                  (GlobalValue::LinkageTypes)Linkage, Name,
                                  Aliasee, /*Parent*/ nullptr));
   GA->setThreadLocalMode(TLM);
   GA->setVisibility((GlobalValue::VisibilityTypes)Visibility);
   GA->setDLLStorageClass((GlobalValue::DLLStorageClassTypes)DLLStorageClass);
   GA->setUnnamedAddr(UnnamedAddr);
   maybeSetDSOLocal(DSOLocal, *GA);
 
+  while (Lex.getKind() == lltok::comma) {
+    Lex.Lex();
+
+    if (Lex.getKind() == lltok::kw_partition) {
+      Lex.Lex();
+      GA->setPartition(Lex.getStrVal());
+      if (ParseToken(lltok::StringConstant, "expected partition string"))
+        return true;
+    } else {
+      return TokError("unknown alias or ifunc property!");
+    }
+  }
+
   if (Name.empty())
     NumberedVals.push_back(GA.get());
 
   if (GVal) {
     // Verify that types agree.
     if (GVal->getType() != GA->getType())
       return Error(
           ExplicitTypeLoc,
bool LLParser::ParseGlobal(const std::string &Name, LocTy NameLoc,
   while (Lex.getKind() == lltok::comma) {
     Lex.Lex();
 
     if (Lex.getKind() == lltok::kw_section) {
       Lex.Lex();
       GV->setSection(Lex.getStrVal());
       if (ParseToken(lltok::StringConstant, "expected global section string"))
         return true;
+    } else if (Lex.getKind() == lltok::kw_partition) {
+      Lex.Lex();
+      GV->setPartition(Lex.getStrVal());
+      if (ParseToken(lltok::StringConstant, "expected partition string"))
+        return true;
     } else if (Lex.getKind() == lltok::kw_align) {
       unsigned Alignment;
       if (ParseOptionalAlignment(Alignment)) return true;
       GV->setAlignment(Alignment);
     } else if (Lex.getKind() == lltok::MetadataVar) {
       if (ParseGlobalObjectMetadataAttachment(*GV))
         return true;
     } else {
if (Lex.getKind() != lltok::lparen)
     return TokError("expected '(' in function argument list");
 
   SmallVector<ArgInfo, 8> ArgList;
   bool isVarArg;
   AttrBuilder FuncAttrs;
   std::vector<unsigned> FwdRefAttrGrps;
   LocTy BuiltinLoc;
   std::string Section;
+  std::string Partition;
   unsigned Alignment;
   std::string GC;
   GlobalValue::UnnamedAddr UnnamedAddr = GlobalValue::UnnamedAddr::None;
   unsigned AddrSpace = 0;
   Constant *Prefix = nullptr;
   Constant *Prologue = nullptr;
   Constant *PersonalityFn = nullptr;
   Comdat *C;
 
   if (ParseArgumentList(ArgList, isVarArg) ||
       ParseOptionalUnnamedAddr(UnnamedAddr) ||
       ParseOptionalProgramAddrSpace(AddrSpace) ||
       ParseFnAttributeValuePairs(FuncAttrs, FwdRefAttrGrps, false,
                                  BuiltinLoc) ||
       (EatIfPresent(lltok::kw_section) &&
        ParseStringConstant(Section)) ||
+      (EatIfPresent(lltok::kw_partition) &&
+       ParseStringConstant(Partition)) ||
       parseOptionalComdat(FunctionName, C) ||
       ParseOptionalAlignment(Alignment) ||
       (EatIfPresent(lltok::kw_gc) &&
        ParseStringConstant(GC)) ||
       (EatIfPresent(lltok::kw_prefix) &&
        ParseGlobalTypeAndValue(Prefix)) ||
       (EatIfPresent(lltok::kw_prologue) &&
        ParseGlobalTypeAndValue(Prologue)) ||
bool LLParser::ParseFunctionHeader(Function *&Fn, bool isDefine) {
   maybeSetDSOLocal(DSOLocal, *Fn);
   Fn->setVisibility((GlobalValue::VisibilityTypes)Visibility);
   Fn->setDLLStorageClass((GlobalValue::DLLStorageClassTypes)DLLStorageClass);
   Fn->setCallingConv(CC);
   Fn->setAttributes(PAL);
   Fn->setUnnamedAddr(UnnamedAddr);
   Fn->setAlignment(Alignment);
   Fn->setSection(Section);
+  Fn->setPartition(Partition);
   Fn->setComdat(C);
   Fn->setPersonalityFn(PersonalityFn);
   if (!GC.empty()) Fn->setGC(GC);
   Fn->setPrefixData(Prefix);
   Fn->setPrologueData(Prologue);
   ForwardRefAttrGroups[Fn] = FwdRefAttrGrps;
 
   // Add all of the arguments we parsed to the function.

llvm/lib/AsmParser/LLToken.h

enum Kind {
   kw_nuw,
   kw_nsw,
   kw_exact,
   kw_inbounds,
   kw_inrange,
   kw_align,
   kw_addrspace,
   kw_section,
+  kw_partition,
   kw_alias,
   kw_ifunc,
   kw_module,
   kw_asm,
   kw_sideeffect,
   kw_alignstack,
   kw_inteldialect,
   kw_gc,

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

if (Record.size() > 12) {
     NewGV->setAttributes(AS);
   }
 
   if (Record.size() > 13) {
     NewGV->setDSOLocal(getDecodedDSOLocal(Record[13]));
   }
   inferDSOLocal(NewGV);
 
+  if (Record.size() > 15)
+    NewGV->setPartition(StringRef(Strtab.data() + Record[14], Record[15]));
+
   return Error::success();
 }
 
 Error BitcodeReader::parseFunctionRecord(ArrayRef<uint64_t> Record) {
   // v1: [type, callingconv, isproto, linkage, paramattr, alignment, section,
   // visibility, gc, unnamed_addr, prologuedata, dllstorageclass, comdat,
   // prefixdata,  personalityfn, preemption specifier, addrspace] (name in VST)
   // v2: [strtab_offset, strtab_size, v1]
Error BitcodeReader::parseFunctionRecord(ArrayRef<uint64_t> Record) {
   if (Record.size() > 14 && Record[14] != 0)
     FunctionPersonalityFns.push_back(std::make_pair(Func, Record[14] - 1));
 
   if (Record.size() > 15) {
     Func->setDSOLocal(getDecodedDSOLocal(Record[15]));
   }
   inferDSOLocal(Func);
 
+  if (Record.size() > 18)
+    Func->setPartition(StringRef(Strtab.data() + Record[17], Record[18]));
+
   ValueList.push_back(Func);
 
   // If this is a function with a body, remember the prototype we are
   // creating now, so that we can match up the body with them later.
   if (!isProto) {
     Func->setIsMaterializable(true);
     FunctionsWithBodies.push_back(Func);
     DeferredFunctionInfo[Func] = 0;
if (OpNum != Record.size())
       NewGA->setThreadLocalMode(getDecodedThreadLocalMode(Record[OpNum++]));
     if (OpNum != Record.size())
       NewGA->setUnnamedAddr(getDecodedUnnamedAddrType(Record[OpNum++]));
   }
   if (OpNum != Record.size())
     NewGA->setDSOLocal(getDecodedDSOLocal(Record[OpNum++]));
   inferDSOLocal(NewGA);
 
+  if (OpNum + 1 < Record.size()) {
+    NewGA->setPartition(
+        StringRef(Strtab.data() + Record[OpNum], Record[OpNum + 1]));
+    OpNum += 2;
+  }
+
   ValueList.push_back(NewGA);
   IndirectSymbolInits.push_back(std::make_pair(NewGA, Val));
   return Error::success();
 }
 
 Error BitcodeReader::parseModule(uint64_t ResumeBit,
                                  bool ShouldLazyLoadMetadata) {
   if (ResumeBit)

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

for (const GlobalVariable &GV : M.globals()) {
     Vals.push_back(GV.hasSection() ? SectionMap[GV.getSection()] : 0);
     if (GV.isThreadLocal() ||
         GV.getVisibility() != GlobalValue::DefaultVisibility ||
         GV.getUnnamedAddr() != GlobalValue::UnnamedAddr::None ||
         GV.isExternallyInitialized() ||
         GV.getDLLStorageClass() != GlobalValue::DefaultStorageClass ||
         GV.hasComdat() ||
         GV.hasAttributes() ||
-        GV.isDSOLocal()) {
+        GV.isDSOLocal() ||
+        GV.hasPartition()) {
       Vals.push_back(getEncodedVisibility(GV));
       Vals.push_back(getEncodedThreadLocalMode(GV));
       Vals.push_back(getEncodedUnnamedAddr(GV));
       Vals.push_back(GV.isExternallyInitialized());
       Vals.push_back(getEncodedDLLStorageClass(GV));
       Vals.push_back(GV.hasComdat() ? VE.getComdatID(GV.getComdat()) : 0);
 
       auto AL = GV.getAttributesAsList(AttributeList::FunctionIndex);
       Vals.push_back(VE.getAttributeListID(AL));
 
       Vals.push_back(GV.isDSOLocal());
+      Vals.push_back(addToStrtab(GV.getPartition()));
+      Vals.push_back(GV.getPartition().size());
     } else {
       AbbrevToUse = SimpleGVarAbbrev;
     }
 
     Stream.EmitRecord(bitc::MODULE_CODE_GLOBALVAR, Vals, AbbrevToUse);
     Vals.clear();
   }
 
for (const Function &F : M) {
     Vals.push_back(F.hasComdat() ? VE.getComdatID(F.getComdat()) : 0);
     Vals.push_back(F.hasPrefixData() ? (VE.getValueID(F.getPrefixData()) + 1)
                                      : 0);
     Vals.push_back(
         F.hasPersonalityFn() ? (VE.getValueID(F.getPersonalityFn()) + 1) : 0);
 
     Vals.push_back(F.isDSOLocal());
     Vals.push_back(F.getAddressSpace());
+    Vals.push_back(addToStrtab(F.getPartition()));
+    Vals.push_back(F.getPartition().size());
 
     unsigned AbbrevToUse = 0;
     Stream.EmitRecord(bitc::MODULE_CODE_FUNCTION, Vals, AbbrevToUse);
     Vals.clear();
   }
 
   // Emit the alias information.
   for (const GlobalAlias &A : M.aliases()) {
     // ALIAS: [strtab offset, strtab size, alias type, aliasee val#, linkage,
     //         visibility, dllstorageclass, threadlocal, unnamed_addr,
     //         DSO_Local]
     Vals.push_back(addToStrtab(A.getName()));
     Vals.push_back(A.getName().size());
     Vals.push_back(VE.getTypeID(A.getValueType()));
     Vals.push_back(A.getType()->getAddressSpace());
     Vals.push_back(VE.getValueID(A.getAliasee()));
     Vals.push_back(getEncodedLinkage(A));
     Vals.push_back(getEncodedVisibility(A));
     Vals.push_back(getEncodedDLLStorageClass(A));
     Vals.push_back(getEncodedThreadLocalMode(A));
     Vals.push_back(getEncodedUnnamedAddr(A));
     Vals.push_back(A.isDSOLocal());
+    Vals.push_back(addToStrtab(A.getPartition()));
+    Vals.push_back(A.getPartition().size());
 
     unsigned AbbrevToUse = 0;
     Stream.EmitRecord(bitc::MODULE_CODE_ALIAS, Vals, AbbrevToUse);
     Vals.clear();
   }
 
   // Emit the ifunc information.
   for (const GlobalIFunc &I : M.ifuncs()) {
     // IFUNC: [strtab offset, strtab size, ifunc type, address space, resolver
     //         val#, linkage, visibility, DSO_Local]
     Vals.push_back(addToStrtab(I.getName()));
     Vals.push_back(I.getName().size());
     Vals.push_back(VE.getTypeID(I.getValueType()));
     Vals.push_back(I.getType()->getAddressSpace());
     Vals.push_back(VE.getValueID(I.getResolver()));
     Vals.push_back(getEncodedLinkage(I));
     Vals.push_back(getEncodedVisibility(I));
     Vals.push_back(I.isDSOLocal());
+    Vals.push_back(addToStrtab(I.getPartition()));
+    Vals.push_back(I.getPartition().size());
     Stream.EmitRecord(bitc::MODULE_CODE_IFUNC, Vals);
     Vals.clear();
   }
 
   writeValueSymbolTableForwardDecl();
 }
 
 static uint64_t getOptimizationFlags(const Value *V) {

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

if (TM.Options.EmitAddrsig) {
     OutStreamer->EmitAddrsig();
     for (const GlobalValue &GV : M.global_values())
       if (!GV.use_empty() && !GV.isThreadLocal() &&
           !GV.hasDLLImportStorageClass() && !GV.getName().startswith("llvm.") &&
           !GV.hasAtLeastLocalUnnamedAddr())
         OutStreamer->EmitAddrsigSym(getSymbol(&GV));
   }
 
+  // Emit symbol partition specifications (ELF only).
+  if (TM.getTargetTriple().isOSBinFormatELF()) {
+    unsigned UniqueID = 0;
+    for (const GlobalValue &GV : M.global_values()) {
+      if (!GV.hasPartition() || GV.isDeclarationForLinker() ||
+          GV.getVisibility() != GlobalValue::DefaultVisibility)
+        continue;
+
+      OutStreamer->SwitchSection(OutContext.getELFSection(
+          ".llvm_sympart", ELF::SHT_LLVM_SYMPART, 0, 0, "", ++UniqueID));
+      OutStreamer->EmitBytes(GV.getPartition());
+      OutStreamer->EmitZeros(1);
+      OutStreamer->EmitValue(
+          MCSymbolRefExpr::create(getSymbol(&GV), OutContext),
+          MAI->getCodePointerSize());
+    }
+  }
+
   // Allow the target to emit any magic that it wants at the end of the file,
   // after everything else has gone out.
   EmitEndOfAsmFile(M);
 
   MMI = nullptr;
 
   OutStreamer->Finish();
   OutStreamer->reset();

llvm/lib/IR/AsmWriter.cpp

if (GV->hasInitializer()) {
     writeOperand(GV->getInitializer(), false);
   }
 
   if (GV->hasSection()) {
     Out << ", section \"";
     printEscapedString(GV->getSection(), Out);
     Out << '"';
   }
+  if (GV->hasPartition()) {
+    Out << ", partition \"";
+    printEscapedString(GV->getPartition(), Out);
+    Out << '"';
+  }
+
   maybePrintComdat(Out, *GV);
   if (GV->getAlignment())
     Out << ", align " << GV->getAlignment();
 
   SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
   GV->getAllMetadata(MDs);
   printMetadataAttachments(MDs, ", ");
 
void AssemblyWriter::printIndirectSymbol(const GlobalIndirectSymbol *GIS) {
 
   if (!IS) {
     TypePrinter.print(GIS->getType(), Out);
     Out << " <<NULL ALIASEE>>";
   } else {
     writeOperand(IS, !isa<ConstantExpr>(IS));
   }
 
+  if (GIS->hasPartition()) {
+    Out << ", partition \"";
+    printEscapedString(GIS->getPartition(), Out);
+    Out << '"';
+  }
+
   printInfoComment(*GIS);
   Out << '\n';
 }
 
 void AssemblyWriter::printComdat(const Comdat *C) {
   C->print(Out);
 }
 
if (F->getAddressSpace() != 0 || !Mod ||
     Out << " addrspace(" << F->getAddressSpace() << ")";
   if (Attrs.hasAttributes(AttributeList::FunctionIndex))
     Out << " #" << Machine.getAttributeGroupSlot(Attrs.getFnAttributes());
   if (F->hasSection()) {
     Out << " section \"";
     printEscapedString(F->getSection(), Out);
     Out << '"';
   }
+  if (F->hasPartition()) {
+    Out << " partition \"";
+    printEscapedString(F->getPartition(), Out);
+    Out << '"';
+  }
   maybePrintComdat(Out, *F);
   if (F->getAlignment())
     Out << " align " << F->getAlignment();
   if (F->hasGC())
     Out << " gc \"" << F->getGC() << '"';
   if (F->hasPrefixData()) {
     Out << " prefix ";
     writeOperand(F->getPrefixData(), true);

llvm/lib/IR/Globals.cpp

 
 /// copyAttributesFrom - copy all additional attributes (those not needed to
 /// create a GlobalValue) from the GlobalValue Src to this one.
 void GlobalValue::copyAttributesFrom(const GlobalValue *Src) {
   setVisibility(Src->getVisibility());
   setUnnamedAddr(Src->getUnnamedAddr());
   setDLLStorageClass(Src->getDLLStorageClass());
   setDSOLocal(Src->isDSOLocal());
+  setPartition(Src->getPartition());
 }
 
 void GlobalValue::removeFromParent() {
   switch (getValueID()) {
 #define HANDLE_GLOBAL_VALUE(NAME)                                              \
   case Value::NAME##Val:                                                       \
     return static_cast<NAME *>(this)->removeFromParent();
 #include "llvm/IR/Value.def"
if (auto *GA = dyn_cast<GlobalAlias>(this)) {
     return nullptr;
   }
   // ifunc and its resolver are separate things so don't use resolver comdat.
   if (isa<GlobalIFunc>(this))
     return nullptr;
   return cast<GlobalObject>(this)->getComdat();
 }
 
+StringRef GlobalValue::getPartition() const {
+  if (!hasPartition())
+    return "";
+  return getContext().pImpl->GlobalValuePartitions[this];
+}
+
+void GlobalValue::setPartition(StringRef S) {
+  // Do nothing if we're clearing the partition and it is already empty.
+  if (!hasPartition() && S.empty())
+    return;
+
+  // Get or create a stable partition name string and put it in the table in the
+  // context.
+  if (!S.empty())
+    S = getContext().pImpl->Saver.save(S);
+  getContext().pImpl->GlobalValuePartitions[this] = S;
+
+  // Update the HasPartition field. Setting the partition to the empty string
+  // means this global no longer has a partition.
+  HasPartition = !S.empty();
+}
+
 StringRef GlobalObject::getSectionImpl() const {
   assert(hasSection());
   return getContext().pImpl->GlobalObjectSections[this];
 }
 
 void GlobalObject::setSection(StringRef S) {
   // Do nothing if we're clearing the section and it is already empty.
   if (!hasSection() && S.empty())
     return;
 
   // Get or create a stable section name string and put it in the table in the
   // context.
-  if (!S.empty()) {
-    S = getContext().pImpl->SectionStrings.insert(S).first->first();
-  }
+  if (!S.empty())
+    S = getContext().pImpl->Saver.save(S);
   getContext().pImpl->GlobalObjectSections[this] = S;
 
   // Update the HasSectionHashEntryBit. Setting the section to the empty string
   // means this global no longer has a section.
   setGlobalObjectFlag(HasSectionHashEntryBit, !S.empty());
 }
 
 bool GlobalValue::isDeclaration() const {

llvm/lib/IR/LLVMContextImpl.h

 #include "llvm/ADT/FoldingSet.h"
 #include "llvm/ADT/Hashing.h"
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/ADT/StringRef.h"
-#include "llvm/ADT/StringSet.h"
 #include "llvm/BinaryFormat/Dwarf.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DebugInfoMetadata.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/IR/RemarkStreamer.h"
 #include "llvm/IR/TrackingMDRef.h"
 #include "llvm/Support/Allocator.h"
 #include "llvm/Support/Casting.h"
+#include "llvm/Support/StringSaver.h"
 #include "llvm/Support/YAMLTraits.h"
 #include <algorithm>
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
 #include <memory>
 #include <string>
 #include <utility>
#include "llvm/IR/Metadata.def"
 
   std::unique_ptr<ConstantTokenNone> TheNoneToken;
 
   // Basic type instances.
   Type VoidTy, LabelTy, HalfTy, FloatTy, DoubleTy, MetadataTy, TokenTy;
   Type X86_FP80Ty, FP128Ty, PPC_FP128Ty, X86_MMXTy;
   IntegerType Int1Ty, Int8Ty, Int16Ty, Int32Ty, Int64Ty, Int128Ty;
 
-  /// TypeAllocator - All dynamically allocated types are allocated from this.
-  /// They live forever until the context is torn down.
-  BumpPtrAllocator TypeAllocator;
+  BumpPtrAllocator Alloc;
+  UniqueStringSaver Saver{Alloc};
 
   DenseMap<unsigned, IntegerType*> IntegerTypes;
 
   using FunctionTypeSet = DenseSet<FunctionType *, FunctionTypeKeyInfo>;
   FunctionTypeSet FunctionTypes;
   using StructTypeSet = DenseSet<StructType *, AnonStructTypeKeyInfo>;
   StructTypeSet AnonStructTypes;
   StringMap<StructType*> NamedStructTypes;
#include "llvm/IR/Metadata.def"
   DenseMap<const Instruction *, MDAttachmentMap> InstructionMetadata;
 
   /// Collection of per-GlobalObject metadata used in this context.
   DenseMap<const GlobalObject *, MDGlobalAttachmentMap> GlobalObjectMetadata;
 
   /// Collection of per-GlobalObject sections used in this context.
   DenseMap<const GlobalObject *, StringRef> GlobalObjectSections;
 
-  /// Stable collection of section strings.
-  StringSet<> SectionStrings;
+  /// Collection of per-GlobalValue partitions used in this context.
+  DenseMap<const GlobalValue *, StringRef> GlobalValuePartitions;
 
   /// DiscriminatorTable - This table maps file:line locations to an
   /// integer representing the next DWARF path discriminator to assign to
   /// instructions in different blocks at the same location.
   DenseMap<std::pair<const char *, unsigned>, unsigned> DiscriminatorTable;
 
   int getOrAddScopeRecordIdxEntry(MDNode *N, int ExistingIdx);
   int getOrAddScopeInlinedAtIdxEntry(MDNode *Scope, MDNode *IA,int ExistingIdx);

llvm/lib/IR/Type.cpp

IntegerType *IntegerType::get(LLVMContext &C, unsigned NumBits) {
   case 128: return cast<IntegerType>(Type::getInt128Ty(C));
   default:
     break;
   }
 
   IntegerType *&Entry = C.pImpl->IntegerTypes[NumBits];
 
   if (!Entry)
-    Entry = new (C.pImpl->TypeAllocator) IntegerType(C, NumBits);
+    Entry = new (C.pImpl->Alloc) IntegerType(C, NumBits);
 
   return Entry;
 }
 
 bool IntegerType::isPowerOf2ByteWidth() const {
   unsigned BitWidth = getBitWidth();
   return (BitWidth > 7) && isPowerOf2_32(BitWidth);
 }
FunctionType *FunctionType::get(Type *ReturnType,
   // and we don't want to perform two lookups (one for checking if existent and
   // one for inserting the newly allocated one), here we instead lookup based on
   // Key and update the reference to the function type in-place to a newly
   // allocated one if not found.
   auto Insertion = pImpl->FunctionTypes.insert_as(nullptr, Key);
   if (Insertion.second) {
     // The function type was not found. Allocate one and update FunctionTypes
     // in-place.
-    FT = (FunctionType *)pImpl->TypeAllocator.Allocate(
+    FT = (FunctionType *)pImpl->Alloc.Allocate(
         sizeof(FunctionType) + sizeof(Type *) * (Params.size() + 1),
         alignof(FunctionType));
     new (FT) FunctionType(ReturnType, Params, isVarArg);
     *Insertion.first = FT;
   } else {
     // The function type was found. Just return it.
     FT = *Insertion.first;
   }
StructType *StructType::get(LLVMContext &Context, ArrayRef<Type*> ETypes,
   // and we don't want to perform two lookups (one for checking if existent and
   // one for inserting the newly allocated one), here we instead lookup based on
   // Key and update the reference to the struct type in-place to a newly
   // allocated one if not found.
   auto Insertion = pImpl->AnonStructTypes.insert_as(nullptr, Key);
   if (Insertion.second) {
     // The struct type was not found. Allocate one and update AnonStructTypes
     // in-place.
-    ST = new (Context.pImpl->TypeAllocator) StructType(Context);
+    ST = new (Context.pImpl->Alloc) StructType(Context);
     ST->setSubclassData(SCDB_IsLiteral);  // Literal struct.
     ST->setBody(ETypes, isPacked);
     *Insertion.first = ST;
   } else {
     // The struct type was found. Just return it.
     ST = *Insertion.first;
   }
 
void StructType::setBody(ArrayRef<Type*> Elements, bool isPacked) {
 
   NumContainedTys = Elements.size();
 
   if (Elements.empty()) {
     ContainedTys = nullptr;
     return;
   }
 
-  ContainedTys = Elements.copy(getContext().pImpl->TypeAllocator).data();
+  ContainedTys = Elements.copy(getContext().pImpl->Alloc).data();
 }
 
 void StructType::setName(StringRef Name) {
   if (Name == getName()) return;
 
   StringMap<StructType *> &SymbolTable = getContext().pImpl->NamedStructTypes;
 
   using EntryTy = StringMap<StructType *>::MapEntryTy;
if (SymbolTableEntry)
     ((EntryTy *)SymbolTableEntry)->Destroy(SymbolTable.getAllocator());
   SymbolTableEntry = &*IterBool.first;
 }
 
 //===----------------------------------------------------------------------===//
 // StructType Helper functions.
 
 StructType *StructType::create(LLVMContext &Context, StringRef Name) {
-  StructType *ST = new (Context.pImpl->TypeAllocator) StructType(Context);
+  StructType *ST = new (Context.pImpl->Alloc) StructType(Context);
   if (!Name.empty())
     ST->setName(Name);
   return ST;
 }
 
 StructType *StructType::get(LLVMContext &Context, bool isPacked) {
   return get(Context, None, isPacked);
 }

 ArrayType *ArrayType::get(Type *ElementType, uint64_t NumElements) {
   assert(isValidElementType(ElementType) && "Invalid type for array element!");
 
   LLVMContextImpl *pImpl = ElementType->getContext().pImpl;
   ArrayType *&Entry =
     pImpl->ArrayTypes[std::make_pair(ElementType, NumElements)];
 
   if (!Entry)
-    Entry = new (pImpl->TypeAllocator) ArrayType(ElementType, NumElements);
+    Entry = new (pImpl->Alloc) ArrayType(ElementType, NumElements);
   return Entry;
 }
 
 bool ArrayType::isValidElementType(Type *ElemTy) {
   return !ElemTy->isVoidTy() && !ElemTy->isLabelTy() &&
          !ElemTy->isMetadataTy() && !ElemTy->isFunctionTy() &&
          !ElemTy->isTokenTy();
 }
assert(isValidElementType(ElementType) && "Element type of a VectorType must "
                                             "be an integer, floating point, or "
                                             "pointer type.");
 
   LLVMContextImpl *pImpl = ElementType->getContext().pImpl;
   VectorType *&Entry = ElementType->getContext().pImpl
     ->VectorTypes[std::make_pair(ElementType, NumElements)];
 
   if (!Entry)
-    Entry = new (pImpl->TypeAllocator) VectorType(ElementType, NumElements);
+    Entry = new (pImpl->Alloc) VectorType(ElementType, NumElements);
   return Entry;
 }
 
 bool VectorType::isValidElementType(Type *ElemTy) {
   return ElemTy->isIntegerTy() || ElemTy->isFloatingPointTy() ||
     ElemTy->isPointerTy();
 }
 
 //===----------------------------------------------------------------------===//
 //                         PointerType Implementation
 //===----------------------------------------------------------------------===//
 
 PointerType *PointerType::get(Type *EltTy, unsigned AddressSpace) {
   assert(EltTy && "Can't get a pointer to <null> type!");
   assert(isValidElementType(EltTy) && "Invalid type for pointer element!");
 
   LLVMContextImpl *CImpl = EltTy->getContext().pImpl;
 
   // Since AddressSpace #0 is the common case, we special case it.
   PointerType *&Entry = AddressSpace == 0 ? CImpl->PointerTypes[EltTy]
      : CImpl->ASPointerTypes[std::make_pair(EltTy, AddressSpace)];
 
   if (!Entry)
-    Entry = new (CImpl->TypeAllocator) PointerType(EltTy, AddressSpace);
+    Entry = new (CImpl->Alloc) PointerType(EltTy, AddressSpace);
   return Entry;
 }
 
 PointerType::PointerType(Type *E, unsigned AddrSpace)
   : Type(E->getContext(), PointerTyID), PointeeTy(E) {
   ContainedTys = &PointeeTy;
   NumContainedTys = 1;
   setSubclassData(AddrSpace);

llvm/lib/MC/MCSectionELF.cpp

else if (Type == ELF::SHT_MIPS_DWARF)
     // any standard symbolic representation of the flag.
     OS << "0x7000001e";
   else if (Type == ELF::SHT_LLVM_ODRTAB)
     OS << "llvm_odrtab";
   else if (Type == ELF::SHT_LLVM_LINKER_OPTIONS)
     OS << "llvm_linker_options";
   else if (Type == ELF::SHT_LLVM_CALL_GRAPH_PROFILE)
     OS << "llvm_call_graph_profile";
+  else if (Type == ELF::SHT_LLVM_SYMPART)
+    OS << "llvm_sympart";
   else
     report_fatal_error("unsupported type 0x" + Twine::utohexstr(Type) +
                        " for section " + getSectionName());
 
   if (EntrySize) {
     assert(Flags & ELF::SHF_MERGE);
     OS << "," << EntrySize;
   }

llvm/test/Bitcode/compatibility.ll

 ; Global Variables -- ExternallyInitialized
 @g.externally_initialized = external externally_initialized global i32
 ; CHECK: @g.externally_initialized = external externally_initialized global i32
 
 ; Global Variables -- section
 @g.section = global i32 0, section "_DATA"
 ; CHECK: @g.section = global i32 0, section "_DATA"
 
+; Global Variables -- partition
[e-mail] = global i32 0, partition "part"
+; CHECK: @g.partition = global i32 0, partition "part"
+
 ; Global Variables -- comdat
 @comdat.any = global i32 0, comdat
 ; CHECK: @comdat.any = global i32 0, comdat
 @comdat.exactmatch = global i32 0, comdat
 ; CHECK: @comdat.exactmatch = global i32 0, comdat
 @comdat.largest = global i32 0, comdat
 ; CHECK: @comdat.largest = global i32 0, comdat
 @comdat.noduplicates = global i32 0, comdat

 ; CHECK: @a.localexec = thread_local(localexec) alias i32, i32* @g.localexec
 
 ; Aliases -- unnamed_addr and local_unnamed_addr
 @a.unnamed_addr = unnamed_addr alias i32, i32* @g.unnamed_addr
 ; CHECK: @a.unnamed_addr = unnamed_addr alias i32, i32* @g.unnamed_addr
 @a.local_unnamed_addr = local_unnamed_addr alias i32, i32* @g.local_unnamed_addr
 ; CHECK: @a.local_unnamed_addr = local_unnamed_addr alias i32, i32* @g.local_unnamed_addr
 
+; Aliases -- partition
+; CHECK: @alias.partition = alias i32, i32* @g.partition, partition "part"
[e-mail] = alias i32, i32* @g.partition, partition "part"
+
 ;; IFunc
 ; Format @<Name> = [Linkage] [Visibility] ifunc <IFuncTy>,
 ;                  <ResolverTy>* @<Resolver>
 
 ; IFunc -- Linkage
 @ifunc.external = external ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.external = ifunc void (), i8* ()* @ifunc_resolver
 @ifunc.private = private ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.private = private ifunc void (), i8* ()* @ifunc_resolver
 @ifunc.internal = internal ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.internal = internal ifunc void (), i8* ()* @ifunc_resolver
 
 ; IFunc -- Visibility
 @ifunc.default = default ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.default = ifunc void (), i8* ()* @ifunc_resolver
 @ifunc.hidden = hidden ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.hidden = hidden ifunc void (), i8* ()* @ifunc_resolver
 @ifunc.protected = protected ifunc void (), i8* ()* @ifunc_resolver
 ; CHECK: @ifunc.protected = protected ifunc void (), i8* ()* @ifunc_resolver
 
+; IFunc -- partition
+; CHECK: @ifunc.partition = ifunc void (), i8* ()* @ifunc_resolver, partition "part"
[e-mail] = ifunc void (), i8* ()* @ifunc_resolver, partition "part"
+
 define i8* @ifunc_resolver() {
 entry:
   ret i8* null
 }
 
 ;; Functions
 ; Format: define [linkage] [visibility] [DLLStorageClass]
 ;         [cconv] [ret attrs]

 declare void @f.inaccessiblemem_or_argmemonly() inaccessiblemem_or_argmemonly
 ; CHECK: declare void @f.inaccessiblemem_or_argmemonly() #34
 declare void @f.strictfp() #35
 
 ; Functions -- section
 declare void @f.section() section "80"
 ; CHECK: declare void @f.section() section "80"
 
+; Functions -- partition
+define void @f.partition() partition "part" {
+; CHECK: define void @f.partition() partition "part"
+  ret void
+}
+
 ; Functions -- comdat
 define void @f.comdat_any() comdat($comdat.any) {
 ; CHECK: define void @f.comdat_any() comdat($comdat.any)
 entry:
   ret void
 }
 define void @f.comdat_exactmatch() comdat($comdat.exactmatch) {
 ; CHECK: define void @f.comdat_exactmatch() comdat($comdat.exactmatch)

llvm/test/CodeGen/X86/partition.ll

+; RUN: llc < %s -mtriple=x86_64-unknown-linux | FileCheck %s
+
+; CHECK: .section .llvm_sympart,"",@llvm_sympart,unique,1
+; CHECK-NEXT: .ascii "part1"
+; CHECK-NEXT: .zero 1
+; CHECK-NEXT: .quad f1
+; CHECK-NEXT: .section .llvm_sympart,"",@llvm_sympart,unique,2
+; CHECK-NEXT: .ascii "part4"
+; CHECK-NEXT: .zero 1
+; CHECK-NEXT: .quad g1
+; CHECK-NEXT: .section .llvm_sympart,"",@llvm_sympart,unique,3
+; CHECK-NEXT: .ascii "part5"
+; CHECK-NEXT: .zero 1
+; CHECK-NEXT: .quad a1
+; CHECK-NEXT: .section .llvm_sympart,"",@llvm_sympart,unique,4
+; CHECK-NEXT: .ascii "part6"
+; CHECK-NEXT: .zero 1
+; CHECK-NEXT: .quad i1
+
+define void @f1() partition "part1" {
+  unreachable
+}
+
+define hidden void @f2() partition "part2" {
+  unreachable
+}
+
+declare void @f3() partition "part3"
+
+@g1 = global i32 0, partition "part4"
+
+@a1 = alias i32, i32* @g1, partition "part5"
+@i1 = ifunc void(), void()* @f1, partition "part6"

llvm/test/Object/X86/irsymtab.ll

 ; RUN: env LLVM_OVERRIDE_PRODUCER=producer opt -o %t %s
 ; RUN: llvm-bcanalyzer -dump -show-binary-blobs %t | FileCheck --check-prefix=BCA %s
 
 ; Same producer, does not require upgrade.
 ; RUN: env LLVM_OVERRIDE_PRODUCER=producer llvm-lto2 dump-symtab %t | FileCheck --check-prefix=SYMTAB %s
 
 ; Different producer, requires upgrade.
 ; RUN: env LLVM_OVERRIDE_PRODUCER=consumer llvm-lto2 dump-symtab %t | FileCheck --check-prefix=SYMTAB %s
 
 ; BCA:      <SYMTAB_BLOCK
 ; Version stored at offset 0.
-; BCA-NEXT:   <BLOB abbrevid=4/> blob data = '\x01\x00\x00\x00\x06\x00\x00\x00\x08\x00\x00\x00D\x00\x00\x00\x01\x00\x00\x00P\x00\x00\x00\x00\x00\x00\x00P\x00\x00\x00\x02\x00\x00\x00\x80\x00\x00\x00\x00\x00\x00\x00\x0E\x00\x00\x00\x18\x00\x00\x00&\x00\x00\x00\x0B\x00\x00\x001\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\xFF\xFF\xFF\xFF\x00$\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\xFF\xFF\xFF\xFF\x08$\x00\x00'
+; BCA-NEXT:   <BLOB abbrevid=4/> blob data = '\x01\x00\x00\x00\x06\x00\x00\x00\x08\x00\x00\x00D\x00\x00\x00\x01\x00\x00\x00P\x00\x00\x00\x00\x00\x00\x00P\x00\x00\x00\x02\x00\x00\x00\x80\x00\x00\x00\x00\x00\x00\x00\x0E\x00\x00\x00\x18\x00\x00\x00&\x00\x00\x00\x0B\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\xFF\xFF\xFF\xFF\x00$\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\x03\x00\x00\x00\xFF\xFF\xFF\xFF\x08$\x00\x00'
 ; BCA-NEXT: </SYMTAB_BLOCK>
 ; BCA-NEXT: <STRTAB_BLOCK
 ; BCA-NEXT:   <BLOB abbrevid=4/> blob data = 'foobarproducerx86_64-unknown-linux-gnuirsymtab.ll'
 ; BCA-NEXT: </STRTAB_BLOCK>
 
 ; SYMTAB:      version: 1
 ; SYMTAB-NEXT: producer: producer
 ; SYMTAB-NEXT: target triple: x86_64-unknown-linux-gnu