[Mono AOT] Improve build time on Apple mobile

[kotlarmilos]

Description

The build time for Apple mobile apps has significantly regressed in .NET 9: #110406.

Simple perf tests #110406 (comment) using the MAUI iOS template app on a macOS-arm64 host with an ios-arm64 target indicate a clear performance regression in the Mono AOT compiler:

Mono AOT	.NET 8 SDK	.NET 9 SDK
NET 8	44.837s	46.303s
NET 9	66s	108.209s

A sample profiling run indicates dn_simdhash_ght_try_get_value hotspot within the Mono AOT compilation:

Total number in stack (recursive counted multiple, when >=5):
        1713       dn_simdhash_ght_try_get_value  (in mono-aot-cross) + 208  [0x10001ec9c]  dn-simdhash-specialization.h:482
        1662       mono_metadata_get_generic_inst  (in mono-aot-cross) + 460  [0x100196194]  metadata.c:3467
        1651       mono_metadata_get_canonical_generic_inst  (in mono-aot-cross) + 212  [0x100196324]  metadata.c:3502
        1548       dn_simdhash_ght_equals  (in mono-aot-cross) + 72  [0x100020c60]  dn-simdhash-ght-compatible.c:35
        1488       dn_simdhash_ght_try_get_value_with_hash  (in mono-aot-cross) + 1020  [0x10001f0a8]  dn-simdhash-specialization.h:490
        1487       mono_metadata_generic_inst_equal  (in mono-aot-cross) + 52  [0x100193384]  metadata.c:1922
        1294       mono_metadata_inflate_generic_inst  (in mono-aot-cross) + 276  [0x1001970d4]  metadata.c:3661
        1227       mono_generic_inst_equal_full  (in mono-aot-cross) + 348  [0x1001934ec]  metadata.c:1910
        1181       dn_simdhash_ght_try_get_value_with_hash  (in mono-aot-cross) + 0  [0x10001ecac]  dn-simdhash-specialization.h:490
        1090       do_mono_metadata_type_equal  (in mono-aot-cross) + 0  [0x10019bac0]  metadata.c:5854


Sort by top of stack, same collapsed (when >= 5):
        __psynch_cvwait  (in libsystem_kernel.dylib)        213507
        mono_generic_inst_equal_full  (in mono-aot-cross)        56454
        do_mono_metadata_type_equal  (in mono-aot-cross)        42611
        dn_simdhash_ght_try_get_value_with_hash  (in mono-aot-cross)        30071
        mono_metadata_generic_param_equal_internal  (in mono-aot-cross)        17027
        dn_simdhash_ght_equals  (in mono-aot-cross)        8150
        mono_metadata_generic_inst_equal  (in mono-aot-cross)        7983
        m_type_data_get_generic_param  (in mono-aot-cross)        5280
        mono_local_regalloc  (in mono-aot-cross)        2434
        m_type_data_get_generic_param_unchecked  (in mono-aot-cross)        2118

Tasks

• [Android][Apple] Setup build-time performance measurements for mobile platforms in the perf lab #113073
• [Mono AOT] Improve dn_simdhash lookup on arm64 #113074
• [Mono AOT] Emit dedup generic types directly into source assemblies #113075
• [Mono AOT] Skip compiling methods from the METHODSPEC table that are trimmed away #113076

[kotlarmilos]

Build time measurements: .NET 9 MAUI template app in fullAOT mode

Dedup	Total AOTCompile (s)	Total JIT time (ms)	Total generation time (ms)
enabled	74.464	22327	12716
disabled	35.351	106053	25073

Since JIT time and generation time are only parts of the AOT compiler that run in parallel, their sum should not match the AOTCompile time, but there is a linear dependency.

From these results, there is a noticeable build time difference when dedup is enabled. However, due to the sequential nature of aot-instances.dll compilation, the potential build time gains are not fully utilized.

[kotlarmilos]

Number of dedup-ed instances: .NET 9 MAUI template app in fullAOT mode

System.Private.CoreLib.dll: 468974
System.Collections.Immutable.dll: 14157
System.Linq.Parallel.dll: 7959
System.Text.Json.dll: 6800
System.Collections.dll: 4024
System.Text.RegularExpressions.dll: 3428
System.Collections.Concurrent.dll: 2733
System.Memory.dll: 2642
System.Linq.dll: 2499
System.Data.Common.dll: 2363
System.Linq.Expressions.dll: 1794
System.Private.DataContractSerialization.dll: 1734
System.Diagnostics.DiagnosticSource.dll: 1055
System.Security.Cryptography.dll: 883
System.Private.Xml.dll: 862
System.Threading.Channels.dll: 302
System.Threading.Tasks.Parallel.dll: 297
System.Runtime.Numerics.dll: 252
System.Net.Http.dll: 235
System.Net.Primitives.dll: 217
System.Reflection.Metadata.dll: 175
Microsoft.CSharp.dll: 124
System.Net.Security.dll: 122
Microsoft.Win32.Registry.dll: 102
System.IO.FileSystem.AccessControl.dll: 102
System.IO.FileSystem.Watcher.dll: 102
System.IO.Pipes.AccessControl.dll: 102
System.IO.Pipelines.dll: 102
System.Net.Quic.dll: 102
System.Runtime.InteropServices.JavaScript.dll: 102
System.Security.AccessControl.dll: 102
System.Security.Principal.Windows.dll: 102
System.Text.Encoding.CodePages.dll: 86
System.Formats.Tar.dll: 77
System.Threading.Tasks.Dataflow.dll: 74
System.Net.Sockets.dll: 72
System.Private.Uri.dll: 70
System.Net.WebSockets.dll: 62
System.ComponentModel.Annotations.dll: 54
System.ComponentModel.TypeConverter.dll: 54
System.Net.NameResolution.dll: 54
System.Net.HttpListener.dll: 54
System.Private.Xml.Linq.dll: 46
System.Net.Http.Json.dll: 45
System.Diagnostics.Process.dll: 36
System.IO.Compression.Brotli.dll: 36
System.Diagnostics.TextWriterTraceListener.dll: 33
System.Diagnostics.TraceSource.dll: 32
System.Drawing.Primitives.dll: 32
System.Formats.Asn1.dll: 32
Microsoft.VisualBasic.Core.dll: 32
System.IO.Pipes.dll: 32
System.Linq.Queryable.dll: 32
System.Net.Requests.dll: 32
System.Net.WebSockets.Client.dll: 32
System.Reflection.Emit.dll: 32
System.Runtime.Serialization.Formatters.dll: 32
System.Transactions.Local.dll: 32
System.ComponentModel.Primitives.dll: 22
System.Collections.NonGeneric.dll: 18
System.Collections.Specialized.dll: 14
System.Diagnostics.FileVersionInfo.dll: 14
System.IO.Compression.ZipFile.dll: 14
System.IO.FileSystem.DriveInfo.dll: 14
System.IO.IsolatedStorage.dll: 14
System.IO.MemoryMappedFiles.dll: 14
System.IO.Compression.dll: 14
System.Net.NetworkInformation.dll: 14
System.Net.WebHeaderCollection.dll: 14
System.Net.Mail.dll: 14
System.ObjectModel.dll: 14
System.Reflection.DispatchProxy.dll: 14
System.Runtime.InteropServices.dll: 14
System.Security.Claims.dll: 14
System.Threading.dll: 14
System.Web.HttpUtility.dll: 5

From these results, System.Private.CoreLib.dll contains the most dedup-ed instances. By comparing the .NET 8 and .NET 9 builds, there are approx. 25,000 new symbols. The number of symbols is retrieved by using the nm command on the binaries and comparing literal matches.

[ivanpovazan]

Just to add a note on:

We want to validate potential impact of improvements before making changes by:

We would also like to measure app size implications.