Improving ARM64 Performance in .NET 7.0

[kunalspathak]

In .NET 7.0, we will continue our efforts to improve the Arm64 code quality and closing the performance gap with x64. Similar to how we did this in .NET 5 in #35853 , we will continue the trend of tracking all the Arm64 issues in a top level issue.

• Arm64: Better addressing mode for float/double array access Arm64: Better addressing mode for float/double array access #64819
• (kunalspathak) Hoisting the invariant out of multi-level nested loops Hoisting the invariant out of multi-level nested loops #61420
• ARM64: Optimize a % b operation ARM64: Optimize a % b operation #34937
• Arm64: Add support for loop alignment
  • Arm64: Align methods containing loops to 32B Arm64: Align methods containing loops to 32B #59828
  • Loop Alignment support for Arm64 Loop Alignment support for Arm64 #60135
  • Hide 'align' instruction behind jmp Hide 'align' instruction behind jmp #60787
• Arm64: Have CpBlkUnroll and InitBlkUnroll use SIMD registers  #68085
• Arm64: Forward memset/memcpy to CRT implementation  #67326 @a74nh
• Tooling support for Arm64 performance investigation
  • jitdump output not accepted by ARM streamline jitdump output not accepted by ARM streamline #62456
• (@kunalspathak) gcenv.interlocked's Interlocked use full memory barriers even with 8.1 Atomics #67824
• (@mangod9) Arm64: Revisit the heuristics for IO completion poller threads #67266 - Revisit after concurrent queue fix
• (@EgorBo, @kunalspathak) x64 vs ARM64 Microbenchmarks Performance Study Report x64 vs ARM64 Microbenchmarks Performance Study Report #67339
• (@EgorBo) Arm64: Better addressing mode for array access whose elements are accessed byref #67981
• (@a74nh) Conditional comparison and selection instructions Arm64: Generate conditional comparison and selection instructions #55364, Arm64: Use csel and ccmp for conditional moves #67894
• Intrinsify some .NET libraries with Arm64 intrinsics
  • (@a74nh) Optimize System.Text.ASCIIUtility for arm64 using cross-platform intrinsics #41292
  • Optimize System.Buffers for arm64 using cross-platform intrinsics #35033

Moved to Future Work

• Arm64: Evaluate if it is possible to combine subsequent field loads in a single load Arm64: Evaluate if it is possible to combine subsequent field loads in a single load #64815
• [Arm64] Peephole optimization opportunities [Arm64] Peephole optimization opportunities #55365
• [ARM64/Linux] Inefficient conditionals branching #12735
• JIT: Redundant fmov's on arm64 for a simple function #58954
• Enable multi-register intrinsics support for Arm64 #64921
• [LSRA] Add support for allocating consecutive registers #39457
• Arm64: Consider using "DC ZVA" instruction #67244
• Arm64: Environment.ProcessorCount returns wrong value on higher core machine #67180, Closed Account for availability of multiple processor groups on Windows 11+ #68639 to try a different approach (AntonLapounov)
• [ARM64] Performance regression: Utf8Encoding #41699
• API Proposal : Arm TableVectorLookup and TableVectorExtension intrinsics #1277
• (TiHan) Arm64: In mod operation happening inside the loop, if divisor is an invariant, hoist the divisor checks Arm64: In mod operation happening inside the loop, if divisor is an invariant, hoist the divisor checks #64795
• (@EgorBo) Optimize jump stubs on arm64 #62302
• (@TIHan) Double constants usage in a loop can be CSEed Double constants usage in a loop can be CSEed #35257

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

In .NET 7.0, we will continue our efforts to improve the Arm64 code quality and closing the performance gap with x64. Similar to how we did this in .NET 5 in #35853 , we will continue the trend of tracking all the Arm64 issues in a top level issue.

• Arm64: Better addressing mode for float/double array access Arm64: Better addressing mode for float/double array access #64819
• Arm64: Evaluate if it is possible to combine subsequent field loads in a single load Arm64: Evaluate if it is possible to combine subsequent field loads in a single load #64815
• Arm64: In mod operation happening inside the loop, if divisor is an invariant, hoist the divisor checks Arm64: In mod operation happening inside the loop, if divisor is an invariant, hoist the divisor checks #64795
• Double constants usage in a loop can be CSEed Double constants usage in a loop can be CSEed #35257
• Hoisting the invariant out of multi-level nested loops Hoisting the invariant out of multi-level nested loops #61420
• ARM64: Optimize a % b operation ARM64: Optimize a % b operation #34937
• ARM64: Optimize a % b operation ARM64: Optimize a % b operation #34937
• [Arm64] Peephole optimization opportunities [Arm64] Peephole optimization opportunities #55365
• Arm64: Add support for loop alignment
  • Arm64: Align methods containing loops to 32B Arm64: Align methods containing loops to 32B #59828
  • Loop Alignment support for Arm64 Loop Alignment support for Arm64 #60135
  • Hide 'align' instruction behind jmp Hide 'align' instruction behind jmp #60787
• Tooling support for Arm64 performance investigation
  • jitdump output not accepted by ARM streamline jitdump output not accepted by ARM streamline #62456

Author:	kunalspathak
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged, User Story
Milestone:	-

[kunalspathak]

@dotnet/jit-contrib

[adamsitnik]

Based on #67339 I think it would be good to add #62302 to this list.

[kunalspathak]

Based on #67339 I think it would be good to add #62302 to this list.

Done. Thanks for preparing the report.

[kunalspathak]

.NET 7 items:

Issue	Owner	ETA	Doable in .NET 7
Arm64: Use 8.1 atomics #67824	@kpathak	June' 22	Yes
Arm64: Align methods containing loops to 32B #59828	@kpathak	DONE
Loop Alignment support for Arm64 #60135	@kpathak	DONE
Hide 'align' instruction behind jmp #60787	@kpathak	DONE
Arm64: Better addressing mode for float/double array access #64819	@EgorBo	DONE
Correctly get the last level cache size used by GC #60166	@mangod9	DONE
The thread pool's global queue doesn't scale well on machines with a large processor count #67845	@mangod9	DONE	Yes
Equivalent thread pool change in Kestrel #67845	@sebastienros	Done
Arm64: Environment.ProcessorCount returns wrong value on higher core machine #67180 (WIP: #68639)	@mangod9	June' 22	?
Arm64: Revisit the heuristics for IO completion poller threads #67266	@mangod9	June' 22	?
Optimize jump stubs on arm64 #62302	@EgorBo	June' 22	?
ARM64: Optimize a % b operation #34937	@TIHan	DONE
Double constants usage in a loop can be CSEed #35257 (WIP)	@TIHan	June' 22	?
x64 vs ARM64 Microbenchmarks Performance Study Report #67339	@EgorBo, @kpathak	June' 22	Yes
Arm64: Better addressing mode for array access whose elements are accessed byref #67981	@EgorBo	June' 22	Yes
Arm64: Forward memset/memcpy to CRT implementation #67326	@kpathak	DONE
Arm64: Have CpBlkUnroll and InitBlkUnroll use SIMD registers #68085	@kpathak	DONE
Hoisting the invariant out of multi-level nested loops #61420	@kpathak, @BruceForstall	Done
Arm64: Generate conditional comparison and selection instructions #55364	@a74nh	WIP: #67894	Yes
Optimize System.Text.ASCIIUtility for arm64 using cross-platform intrinsics #41292	@a74nh	Done
Optimize System.Buffers for arm64 using cross-platform intrinsics #35033	@a74nh	Done

Stretch goals:

Issue	Owner	ETA
[Arm64] Peephole optimization opportunities #55365	@a74nh	TBD
[LSRA] Add support for allocating consecutive registers #39457	@kpathak	Future
Enable multi-register intrinsics support for Arm64 #64921	@BruceForstall	Future
API Proposal : Arm TableVectorLookup and TableVectorExtension intrinsics #1277	@a74nh	Future
jitdump output not accepted by ARM streamline #62456	@RobertHenry6bev	Future
Optimize set_brick code in GC	@Maoni0	TBD
[ARM64] Performance regression: Utf8Encoding #41699	@a74nh	TBD
[ARM64/Linux] Inefficient conditionals branching #12735	@a74nh	TBD
JIT: Redundant fmov's on arm64 for a simple function #58954	@a74nh	TBD
Arm64: Consider using "DC ZVA" instruction #67244	@kpathak, @a74nh	TBD
Review the multi-op instruction usage for Arm64 #68028	@TIHan	Future
Arm64: Evaluate if it is possible to combine subsequent field loads in a single load #64815 (lower priority)	TBD	TBD
Arm64: In mod operation happening inside the loop, if divisor is an invariant, hoist the divisor checks #64795	@TIHan	.NET 8

[JulieLeeMSFT]

#68028

[kunalspathak]

#68028

Included in the table above.

[a74nh]

Optimize System.Text.ASCIIUtility for arm64 using cross-platform intrinsics
Issue: #41292
PR: #70080 and #71637
Approach taken:
The existing Sse2/Sse41 implementation was moved to use the Vector128 API.
Arm64 was switched to use the Vector128 implementation instead of the vector generic version.
Where required for performance, the Sse2/Sse41/AdvSimd APIs were used.
Impact:
Small improvement in the relevant microbenchmarks on Arm64.
These changes were not significant enough to be picked up by the Performanceautofiler post merge.
Performance gain was small due to the vector generic version being good/simple.
Use of the Vector128 API helps to reduce code debt.
Follow on work:
None.

Optimize System.Buffers for arm64 using cross-platform intrinsics
Issue: #35033
PR: #70654 and dotnet/performance#2479
Approach taken:
Fixed issue with the microbenchmarks using invalid data.
The existing Ssse3 implementation was moved to use the Vector128 API.
Arm64 was switched to use the Vector128 implementation instead of non-vectorised version.
Where required for performance, the Sse3/AdvSimd APIs were used.
Impact:
Large improvement in the relevant microbenchmarks on Arm64.
Performance improvements detected by the Performanceautofiler:
dotnet/perf-autofiling-issues#6346
dotnet/perf-autofiling-issues#6334
dotnet/perf-autofiling-issues#6340
dotnet/perf-autofiling-issues#6328
dotnet/perf-autofiling-issues#6327
dotnet/perf-autofiling-issues#6326
dotnet/perf-autofiling-issues#6321
Use of the Vector128 API helps to reduce code debt.
Follow on Work:
Once multiple register instructions (such as LD4) have been implemented in Vector128, then further improvements may be possible switching to an implementation based on the NEON version of the Aklomp base64 algorithm.

[kunalspathak]

The only thing remaining from the list is #55364 and it is already set for .NET 7. I will move this to .NET 8 to track remaining work.

[kunalspathak]

Replaced with #77010