Emit GC info into a COMDAT section #83371

MichalStrehovsky · 2023-03-14T04:19:57Z

Saves 3.14% on BasicWebApi. 🤯

We run linker with COMDAT folding enabled. Generating this data into COMDAT foldable section allows linker to deduplicate these.

Cc @dotnet/ilc-contrib

Saves 3.14% on BasicWebApi. 🤯 We run linker with COMDAT folding enabled. Generating this data into COMDAT foldable section allows linker to deduplicate these.

ghost · 2023-03-14T04:20:11Z

Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas
See info in area-owners.md if you want to be subscribed.

Issue Details

Saves 3.14% on BasicWebApi. 🤯

We run linker with COMDAT folding enabled. Generating this data into COMDAT foldable section allows linker to deduplicate these.

Cc @dotnet/ilc-contrib

Author:	MichalStrehovsky
Assignees:	-
Labels:	`area-NativeAOT-coreclr`
Milestone:	-

agocke · 2023-03-14T04:47:10Z

src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/DependencyAnalysis/ObjectWriter.cs

-                ObjectNodeSection section = ObjectNodeSection.XDataSection;
-                if (ShouldShareSymbol(node))
-                    section = GetSharedSection(section, _sb.ToString());
+                ObjectNodeSection section = GetSharedSection(ObjectNodeSection.XDataSection, _sb.ToString());


It looks like LLVM defines a number of different SelectionKinds for COMDAT sections, but I can't see where define any. The LLVM ones look like this:

any
exactMatch
largest
noDuplicates
sameSize

Reason I'm asking is because it sounds like we would probably only want "any" or "noDuplicates"

It appears to do the right thing on Windows, but I haven't checked Linux (hoped I would see it in the CI results).

If it doesn't work on Linux, it's likely because of #77491. I'd back out the non-Windows portion of this change and file a bug to reap this benefit outside Windows in 8.0 (it would be more work than just a one-liner that I randomly tried).

MichalStrehovsky · 2023-03-14T05:43:36Z

Yeah, looks like it doesn't work on Linux.

Before:

/datadisks/disk1/work/A211088F/p/nativeaottest.sh /datadisks/disk1/work/A211088F/w/B4D609BE/e/nativeaot/SmokeTests/HardwareIntrinsics/X64Baseline/ X64Baseline.dll ''
****************************************************
* Size test                                        *
* Size of the executable is   1,916 kB             *
****************************************************

After:

/datadisks/disk1/work/A63308E9/p/nativeaottest.sh /datadisks/disk1/work/A63308E9/w/AEDE099C/e/nativeaot/SmokeTests/HardwareIntrinsics/X64Baseline/ X64Baseline.dll ''
****************************************************
* Size test                                        *
* Size of the executable is   1,916 kB             *
****************************************************

For comparison, on Windows:

call C:\h\w\B28909D9\p\nativeaottest.cmd C:\h\w\B28909D9\w\C78F0A4F\e\nativeaot\SmokeTests\HardwareIntrinsics\X64Baseline\ X64Baseline.dll 
****************************************************
* Size test                                        *
* Size of the executable is   1,835 kB             *
****************************************************

call C:\h\w\B6CE09A2\p\nativeaottest.cmd C:\h\w\B6CE09A2\w\A54E08F8\e\nativeaot\SmokeTests\HardwareIntrinsics\X64Baseline\ X64Baseline.dll 
****************************************************
* Size test                                        *
* Size of the executable is   1,809 kB             *
****************************************************

MichalStrehovsky · 2023-03-14T06:46:36Z

Rolled back the Linux portion and file #83375 for tracking :(.

MichalStrehovsky · 2023-03-14T08:34:06Z

/azp run runtime-extra-platforms

azure-pipelines · 2023-03-14T08:34:29Z

Azure Pipelines successfully started running 1 pipeline(s).

am11 · 2023-03-14T15:48:47Z

Rolled back the Linux portion and file #83375 for tracking :(.

Seems like -Wl,--icf=all and -Wl,--icf=safe etc. do shave off 10.24 KB from hello-world-app on linux-arm64 (using daily build). Unfortunately, bfd linker doesn't support --icf option yet, but LLVM linker -fuse-ld=lld (activated by -p:UseLlvmLinker=true) and Gold linker -fuse-ld=gold support it. Gold linker is better on linux anyway; it's faster, consumes less memory and generates slimmer apps (hello-world-app published with ld.bfd (the default) on linux-arm64: 1918984 bytes while ld.gold: 1853632 bytes). We should probably prefer gold anyway, when it's available.

Introspection in Microsoft.NETCore.Native.Unix.targets will look like this:

  <![CDATA[
  Testing:
  
    if ! cc -fuse-ld=gold -Wl,--icf=all 2>&1 | grep -q unrecognized; then
        echo "yes, gold linker is available"
    fi
  ]]>

  <Exec Command="! &quot;$(CppLinker)&quot; -fuse-ld=gold -Wl,--icf=all 2&gt;&amp;1 | grep -q unrecognized" 
        Condition="'$(UseLlvmLinker)' != 'true' and '$(_targetOS)' == 'linux'"
        IgnoreExitCode="true" 
        StandardOutputImportance="Low">
    <Output TaskParameter="ExitCode" PropertyName="_GoldLinkerSupportedExitCode" />
  </Exec>

  <ItemGroup>
    <LinkerArg Include="-fuse-ld=gold" Condition="'$(_GoldLinkerSupportedExitCode)' == '0'" />
    <LinkerArg Include="-Wl,--icf=all" Condition="'$(_GoldLinkerSupportedExitCode)' == '0' or '$(UseLlvmLinker)' == 'true'" />
  </ItemGroup>

MichalStrehovsky · 2023-03-15T07:57:24Z

Seems like -Wl,--icf=all and -Wl,--icf=safe etc. do shave off 10.24 KB from hello-world-app on linux-arm64 (using daily build)

Thanks for testing it out! I think this is only folding the native code within the executable. We don't generate code into separate sections unless -p:IlcFoldIdenticalMethodBodies=true was specified. This optimization produces non-sensical callstacks so we're unlikely to enable it by default. It's a good addition, but it's unlikely to help more than those 10 kB even for larger apps.

Looks like what we want is -Wl,--icf-data. But it doesn't seem to have an effect, likely due to #77491 (we mess up something in object writing).

Introspection in Microsoft.NETCore.Native.Unix.targets will look like this:

Is this something you'd be willing to contribute? I just tried -fuse-ld=gold on BasicMinimalApi on Ubuntu 18 and just this one switch without anything else reduced the size from 11.5 MB to 10.8 MB (🤯).

am11 · 2023-03-15T09:02:58Z

This optimization produces non-sensical callstacks so we're unlikely to enable it by default. It's a good addition, but it's unlikely to help more than those 10 kB even for larger apps.

Ah, I haven't looked too closely. I found the difference only with -p:UseLlvmLinker=true (fuse-ld=lld which is selected by default in NativeAOT BuildIntegration on FreeBSD). From the binutils ticket:

The lld's safe ICF works with an LLVM feature. Since mid-2018, LLVM emits a
`.llvm_addrsig` section to all object files by default. That section contains
symbol indices whose addresses are taken. Using this table, lld can merge
functions more aggressively than gold can do.

Perhaps we would see more differences with lld if we start emitting this section? Other linkers apparently will optimize using this section in the future and we will automatically opt-in without additional work (bfd: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105625, mold: rui314/mold#484).

Is this something you'd be willing to contribute? I just tried -fuse-ld=gold on BasicMinimalApi on Ubuntu 18 and just this one switch without anything else reduced the size from 11.5 MB to 10.8 MB (🤯).

Sounds great, I will take a look. 🙂

into gcinfofold

MichalStrehovsky · 2023-04-25T04:43:05Z

@dotnet/ilc-contrib could someone have a look? This is a good saving on Windows. We'll have to investigate how to translate this to Linux but that's tracked in #83375. It's likely object writer work around COMDATs.

jkotas

LGTM!

Emit GC info into a COMDAT section

8bfca02

Saves 3.14% on BasicWebApi. 🤯 We run linker with COMDAT folding enabled. Generating this data into COMDAT foldable section allows linker to deduplicate these.

MichalStrehovsky added the area-NativeAOT-coreclr label Mar 14, 2023

ghost assigned MichalStrehovsky Mar 14, 2023

agocke reviewed Mar 14, 2023

View reviewed changes

MichalStrehovsky mentioned this pull request Mar 14, 2023

Investigate COMDAT folding GC info on Linux #83375

Closed

Update ObjectWriter.cs

ada3b00

This was referenced Mar 14, 2023

Tracking issue for CI build timeouts #76454

Closed

[release/6.0] Doublelinklist GC failures on Mono #83245

Closed

[jitstress] HardwareIntrinsics_ro fails with "process cannot access the file" error #83298

Closed

MichalStrehovsky added 2 commits April 25, 2023 12:51

Merge branch 'main' into gcinfofold

011460b

Merge branch 'gcinfofold' of https://github.com/MichalStrehovsky/runtime

b22b0d5

into gcinfofold

jkotas approved these changes Apr 25, 2023

View reviewed changes

MichalStrehovsky merged commit 9a693e0 into dotnet:main Apr 25, 2023

MichalStrehovsky deleted the gcinfofold branch April 25, 2023 06:24

ghost locked as resolved and limited conversation to collaborators May 25, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Emit GC info into a COMDAT section #83371

Emit GC info into a COMDAT section #83371

MichalStrehovsky commented Mar 14, 2023

ghost commented Mar 14, 2023

agocke Mar 14, 2023

agocke Mar 14, 2023

MichalStrehovsky Mar 14, 2023

MichalStrehovsky commented Mar 14, 2023

MichalStrehovsky commented Mar 14, 2023

MichalStrehovsky commented Mar 14, 2023

azure-pipelines bot commented Mar 14, 2023

am11 commented Mar 14, 2023 •

edited

Loading

MichalStrehovsky commented Mar 15, 2023

am11 commented Mar 15, 2023

MichalStrehovsky commented Apr 25, 2023

jkotas left a comment

Emit GC info into a COMDAT section #83371

Emit GC info into a COMDAT section #83371

Conversation

MichalStrehovsky commented Mar 14, 2023

ghost commented Mar 14, 2023

agocke Mar 14, 2023

Choose a reason for hiding this comment

agocke Mar 14, 2023

Choose a reason for hiding this comment

MichalStrehovsky Mar 14, 2023

Choose a reason for hiding this comment

MichalStrehovsky commented Mar 14, 2023

MichalStrehovsky commented Mar 14, 2023

MichalStrehovsky commented Mar 14, 2023

azure-pipelines bot commented Mar 14, 2023

am11 commented Mar 14, 2023 • edited Loading

MichalStrehovsky commented Mar 15, 2023

am11 commented Mar 15, 2023

MichalStrehovsky commented Apr 25, 2023

jkotas left a comment

Choose a reason for hiding this comment

am11 commented Mar 14, 2023 •

edited

Loading