Allow some intrinsics in Tier0 #77357

EgorBo · 2022-10-23T18:08:04Z

Expand some intrinsics in Tier0, e.g. these:

Plaintext-plaintext, Tier0 instrumented:

| load                   | Main            | PR              |         |
| ---------------------- | --------------- | --------------- | ------- |
| First Request (ms)     |             199 |             192 |  -3.52% |
| Requests/sec           |         545,865 |         610,476 | +11.84% |
| Mean latency (ms)      |            4.30 |            3.89 |  -9.53% |
| Latency 50th (ms)      |            4.17 |            3.78 |  -9.35% |
| Latency 75th (ms)      |            6.14 |            5.55 |  -9.61% |
| Latency 90th (ms)      |            7.58 |            6.90 |  -8.97% |

Mainly due to Span.get_Item and typeof() == typeof()

E.g. here is Main's codegen for typeof():

This PR fixes it to just return false

ghost · 2022-10-23T18:08:17Z

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

Issue Details

As it turns out, even extremely hot methods may stuck in Tier0 for several seconds (see #70941 (comment)) pending call-counting stub installation. So we'd better optimize Enum.HasFlag in tier0 too because it allocates and produces a large tree, e.g.:

bool Test(MyEnum e) => e.HasFlag(MyEnum.B);

Current Tier0 codegen:

; Assembly listing for method Program:Test(int):bool:this
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; Tier-0 compilation
; MinOpts code
; rbp based frame
; partially interruptible
G_M000_IG01:                ;; offset=0000H
       55                   push     rbp
       4883EC30             sub      rsp, 48
       488D6C2430           lea      rbp, [rsp+30H]
       33C0                 xor      eax, eax
       488945F8             mov      qword ptr [rbp-08H], rax
       488945F0             mov      qword ptr [rbp-10H], rax
       48894D10             mov      gword ptr [rbp+10H], rcx
       895518               mov      dword ptr [rbp+18H], ed
G_M000_IG02:                ;; offset=001BH
       48B928667DEAFD7F0000 mov      rcx, 0x7FFDEA7D6628
       E8E6C0AB5F           call     CORINFO_HELP_NEWSFAST
       488945F8             mov      gword ptr [rbp-08H], rax
       488B4DF8             mov      rcx, gword ptr [rbp-08H]
       8B4518               mov      eax, dword ptr [rbp+18H]
       894108               mov      dword ptr [rcx+08H], eax
       48B928667DEAFD7F0000 mov      rcx, 0x7FFDEA7D6628
       E8C9C0AB5F           call     CORINFO_HELP_NEWSFAST
       488945F0             mov      gword ptr [rbp-10H], rax
       488B4DF0             mov      rcx, gword ptr [rbp-10H]
       C7410801000000       mov      dword ptr [rcx+08H], 1
       488B4DF8             mov      rcx, gword ptr [rbp-08H]
       488B55F0             mov      rdx, gword ptr [rbp-10H]
       FF15DCF0F4FF         call     [System.Enum:HasFlag(System.Enum):bool:this]
       90                   nop      
G_M000_IG03:                ;; offset=0065H
       4883C430             add      rsp, 48
       5D                   pop      rbp
       C3                   ret      
; Total bytes of code 107

New tier0 codegen:

; Assembly listing for method Program:Test(int):bool:this
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; Tier-0 compilation
; MinOpts code
; rbp based frame
; partially interruptible
G_M57656_IG01:              ;; offset=0000H
       55                   push     rbp
       4883EC20             sub      rsp, 32
       488D6C2420           lea      rbp, [rsp+20H]
       33C0                 xor      eax, eax
       488945F8             mov      qword ptr [rbp-08H], rax
       488945F0             mov      qword ptr [rbp-10H], rax
       48894D10             mov      gword ptr [rbp+10H], rcx
       895518               mov      dword ptr [rbp+18H], edx
G_M57656_IG02:              ;; offset=001BH
       8B4518               mov      eax, dword ptr [rbp+18H]
       8945EC               mov      dword ptr [rbp-14H], eax
       8B45EC               mov      eax, dword ptr [rbp-14H]
       83E001               and      eax, 1
       83F801               cmp      eax, 1
       0F94C0               sete     al
       0FB6C0               movzx    rax, al
G_M57656_IG03:              ;; offset=0030H
       4883C420             add      rsp, 32
       5D                   pop      rbp
       C3                   ret      
; Total bytes of code 54

Also, allow expansion for some other lightweight intrinsics for two reasons:

GT_CALL nodes are not cheap too
We want to avoid unnecessary JIT compilations e.g. why would we need to spend time compiling String.get_Length() + installing a call-counting stub for it when the expansion for it super trivial. Etc.

Author:	EgorBo
Assignees:	EgorBo
Labels:	`area-CodeGen-coreclr`
Milestone:	-

src/coreclr/jit/compiler.h

EgorBo · 2022-10-23T20:36:34Z

Avalonia's version of Enum.HasFlag: https://github.com/AvaloniaUI/Avalonia/blob/28d2765abcac2bd55128611c8a416aef051b6e82/src/Avalonia.Base/EnumExtensions.cs#L12-L42 - it was invented because our Enum.HasFlag is slow and allocates in tier0 (and for net4x support)

AndyAyersMS · 2022-10-23T20:53:41Z

Linking to #9120.

src/coreclr/jit/importer.cpp

src/coreclr/jit/compiler.h

BruceForstall · 2022-11-28T19:47:24Z

@EgorBo No activity on this for a while. Do you want to convert it to Draft until it is ready? Or are you waiting on more reviews?

EgorBo · 2022-11-28T19:48:27Z

@EgorBo No activity on this for a while. Do you want to convert it to Draft until it is ready? Or are you waiting on more reviews?

Yes, it's blocked on the other PR to introduce opt levels.

ghost · 2022-12-28T23:01:53Z

Draft Pull Request was automatically closed for 30 days of inactivity. Please let us know if you'd like to reopen it.

EgorBo · 2023-02-10T19:57:09Z

Reminder for myself to land this.

TechEmpower Tier0:

A lot of time is spent inside Span.get_Item that could be expanded (and is in this PR). Our theory that it's expensive becuase of block counters inside (PGO is enabled) that we don't need for it

EgorBo · 2023-02-11T00:26:07Z

Diffs are negative (MinOpts only): https://dev.azure.com/dnceng-public/public/_build/results?buildId=167821&view=ms.vss-build-web.run-extensions-tab

EgorBo · 2023-02-12T11:02:44Z

😕 (that's for Main, this PR fixes it)

EgorBo · 2023-02-12T11:57:25Z

Plaintext-plaintext, Tier0 instrumented:

| load                   | Main            | PR              |         |
| ---------------------- | --------------- | --------------- | ------- |
| First Request (ms)     |             199 |             192 |  -3.52% |
| Requests/sec           |         545,865 |         610,476 | +11.84% |
| Mean latency (ms)      |            4.30 |            3.89 |  -9.53% |
| Latency 50th (ms)      |            4.17 |            3.78 |  -9.35% |
| Latency 75th (ms)      |            6.14 |            5.55 |  -9.61% |
| Latency 90th (ms)      |            7.58 |            6.90 |  -8.97% |

Mainly due to Span.get_Item and typeof() == typeof()

EgorBo · 2023-02-12T12:16:59Z

Even better numbers for YARP, start time is also decreased!

| application                             | Main        | PR          |         |
| --------------------------------------- | ----------- | ----------- | ------- |
| CPU Usage (%)                           |         100 |          99 |  -1.00% |
| Cores usage (%)                         |       2,794 |       2,777 |  -0.61% |
| Working Set (MB)                        |         460 |         459 |  -0.22% |
| Private Memory (MB)                     |       1,232 |       1,233 |  +0.08% |
| Build Time (ms)                         |       5,127 |       4,346 | -15.23% |
| Start Time (ms)                         |         846 |         802 |  -5.20% |
| Published Size (KB)                     |      90,941 |      90,941 |   0.00% |
| Symbols Size (KB)                       |           0 |           0 |         |
| Max CPU Usage (%)                       |         100 |          99 |  -0.56% |
| Max Working Set (MB)                    |         482 |         481 |  -0.07% |
| Max GC Heap Size (MB)                   |         339 |         339 |  +0.12% |
| Size of committed memory by the GC (MB) |         398 |         403 |  +1.10% |
| Max Number of Gen 0 GCs / sec           |        1.00 |        1.00 |   0.00% |
| Max Number of Gen 1 GCs / sec           |        1.00 |        1.00 |   0.00% |
| Max Number of Gen 2 GCs / sec           |        1.00 |        1.00 |   0.00% |
| Max Time in GC (%)                      |        0.00 |        1.00 |     +∞% |
| Max Gen 0 Size (B)                      |   5,541,728 |   6,963,296 | +25.65% |
| Max Gen 1 Size (B)                      |   9,210,496 |   9,532,536 |  +3.50% |
| Max Gen 2 Size (B)                      |  11,709,272 |  11,713,680 |  +0.04% |
| Max LOH Size (B)                        |      85,072 |      85,072 |   0.00% |
| Max POH Size (B)                        |   1,186,016 |   1,181,896 |  -0.35% |
| Max Allocation Rate (B/sec)             | 113,178,784 | 179,763,992 | +58.83% |
| Max GC Heap Fragmentation               |          34 |          33 |  -3.02% |
| # of Assemblies Loaded                  |         104 |         104 |   0.00% |
| Max Exceptions (#/s)                    |           0 |           0 |         |
| Max Lock Contention (#/s)               |         246 |         304 | +23.58% |
| Max ThreadPool Threads Count            |          48 |          48 |   0.00% |
| Max ThreadPool Queue Length             |         181 |         193 |  +6.63% |
| Max ThreadPool Items (#/s)              |      64,596 |      99,265 | +53.67% |
| Max Active Timers                       |         221 |         212 |  -4.07% |
| IL Jitted (B)                           |     724,634 |     723,985 |  -0.09% |
| Methods Jitted                          |      10,422 |      10,385 |  -0.36% |
| Cores Usage (90th %)                    |       2,780 |       2,750 |  -1.08% |

| load                   | Main        | PR          |         |
| ---------------------- | ----------- |------------ | ------- |
| CPU Usage (%)          |          14 |          19 | +35.71% |
| Cores usage (%)        |         383 |         539 | +40.73% |
| Working Set (MB)       |         125 |         125 |   0.00% |
| Private Memory (MB)    |         217 |         217 |   0.00% |
| Start Time (ms)        |         109 |         106 |  -2.75% |
| First Request (ms)     |         496 |         499 |  +0.60% |
| Requests               |     740,445 |   1,157,557 | +56.33% |
| Latency 50th (ms)      |        5.08 |        3.24 | -36.33% |
| Latency 75th (ms)      |        5.33 |        3.48 | -34.67% |
| Latency 90th (ms)      |        5.69 |        3.76 | -33.88% |
| Latency 95th (ms)      |        5.94 |        3.95 | -33.41% |
| Latency 99th (ms)      |        6.80 |        4.90 | -27.94% |
| Mean latency (ms)      |        5.18 |        3.31 | -36.07% |
| Max latency (ms)       |       78.73 |       91.77 | +16.56% |
| Requests/sec           |      49,478 |      77,391 | +56.42% |
| Requests/sec (max)     |      67,962 |      87,072 | +28.12% |
| Read throughput (MB/s) |        7.15 |       11.19 | +56.35% |
| Cores Usage (90th %)   |         382 |         538 | +40.84% |

EgorBo · 2023-02-12T14:23:35Z

PTAL @AndyAyersMS @dotnet/jit-contrib

src/coreclr/jit/importercalls.cpp

Co-authored-by: Jakob Botsch Nielsen <Jakob.botsch.nielsen@gmail.com>

EgorBo · 2023-02-13T20:07:57Z

ping @AndyAyersMS @jakobbotsch - I need this to land for my next pr 🙂

jakobbotsch

Looks like a reasonable set to expand to me.

AndyAyersMS · 2023-02-14T03:10:57Z

@EgorBo sorry I missed seeing this. Going forward can you assign me as reviewer? Otherwise I may not look in the right places.

Also if you have copies of your evaluation scripts laying around, point me at them.

This reverts commit e1edb16.

BruceForstall · 2023-02-14T08:15:12Z

This caused outerloop regressions: #82078

This reverts commit e1edb16.

Allow some intrinsics in Tier0

c8a1352

ghost assigned EgorBo Oct 23, 2022

dotnet-issue-labeler bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Oct 23, 2022

EgorBo mentioned this pull request Oct 23, 2022

Dynamic PGO startup improvements in NET 8 #76969

Closed

23 tasks

SingleAccretion reviewed Oct 23, 2022

View reviewed changes

src/coreclr/jit/compiler.h Outdated Show resolved Hide resolved

Update compiler.h

78aac50

fix size regressions

e6c4669

jkotas reviewed Oct 24, 2022

View reviewed changes

src/coreclr/jit/importer.cpp Outdated Show resolved Hide resolved

limit to 128 locals

dec24a5

AndyAyersMS reviewed Oct 24, 2022

View reviewed changes

src/coreclr/jit/compiler.h Outdated Show resolved Hide resolved

build-analysis bot mentioned this pull request Oct 24, 2022

Microsoft.DotNet.CoreSetup.Test.HostActivation tests failing in CI #75429

Closed

EgorBo mentioned this pull request Oct 26, 2022

Introduce OptLevel() in jit #77465

Closed

EgorBo marked this pull request as draft November 28, 2022 19:48

ghost closed this Dec 28, 2022

EgorBo reopened this Jan 12, 2023

EgorBo added 4 commits February 10, 2023 21:31

Merge branch 'main' of github.com:dotnet/runtime into allow-tier0-opts

5d292ce

Add span intrinsics

1eb5e10

Clean up

759e39f

clean up

3b5c1c3

build-analysis bot mentioned this pull request Feb 11, 2023

Tracking issue for CI build timeouts #76454

Closed

EgorBo marked this pull request as ready for review February 11, 2023 01:19

EgorBo added 2 commits February 12, 2023 12:11

Merge branch 'main' of github.com:dotnet/runtime into allow-tier0-opts

1d276f9

Clean up, handle typeof() == typeof()

10d3cfe

EgorBo changed the title ~~Allow some intrinsics in Tier0 (e.g. Enum.HasFlag)~~ Allow some intrinsics in Tier0 Feb 12, 2023

Update importercalls.cpp

7f17843

jakobbotsch reviewed Feb 12, 2023

View reviewed changes

src/coreclr/jit/importercalls.cpp Outdated Show resolved Hide resolved

Update src/coreclr/jit/importercalls.cpp

341bce7

Co-authored-by: Jakob Botsch Nielsen <Jakob.botsch.nielsen@gmail.com>

jakobbotsch approved these changes Feb 13, 2023

View reviewed changes

EgorBo merged commit e1edb16 into dotnet:main Feb 13, 2023

EgorBo deleted the allow-tier0-opts branch February 13, 2023 22:46

BruceForstall added a commit to BruceForstall/runtime that referenced this pull request Feb 14, 2023

Revert "Allow some intrinsics in Tier0 (dotnet#77357)"

840e988

This reverts commit e1edb16.

BruceForstall mentioned this pull request Feb 14, 2023

Test failure baseservices\\multidimmarray\\rank1array\\rank1array.cmd #82078

Closed

BruceForstall added a commit that referenced this pull request Feb 14, 2023

Revert "Allow some intrinsics in Tier0 (#77357)" (#82079)

c8aee30

This reverts commit e1edb16.

dotnet locked as resolved and limited conversation to collaborators Mar 16, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Allow some intrinsics in Tier0 #77357

Allow some intrinsics in Tier0 #77357

EgorBo commented Oct 23, 2022 •

edited

ghost commented Oct 23, 2022

EgorBo commented Oct 23, 2022 •

edited

AndyAyersMS commented Oct 23, 2022

BruceForstall commented Nov 28, 2022

EgorBo commented Nov 28, 2022

ghost commented Dec 28, 2022

EgorBo commented Feb 10, 2023 •

edited

EgorBo commented Feb 11, 2023

EgorBo commented Feb 12, 2023 •

edited

EgorBo commented Feb 12, 2023 •

edited

EgorBo commented Feb 12, 2023 •

edited

EgorBo commented Feb 12, 2023

EgorBo commented Feb 13, 2023

jakobbotsch left a comment

AndyAyersMS commented Feb 14, 2023

BruceForstall commented Feb 14, 2023

Allow some intrinsics in Tier0 #77357

Allow some intrinsics in Tier0 #77357

Conversation

EgorBo commented Oct 23, 2022 • edited

Plaintext-plaintext, Tier0 instrumented:

ghost commented Oct 23, 2022

EgorBo commented Oct 23, 2022 • edited

AndyAyersMS commented Oct 23, 2022

BruceForstall commented Nov 28, 2022

EgorBo commented Nov 28, 2022

ghost commented Dec 28, 2022

EgorBo commented Feb 10, 2023 • edited

EgorBo commented Feb 11, 2023

EgorBo commented Feb 12, 2023 • edited

EgorBo commented Feb 12, 2023 • edited

Plaintext-plaintext, Tier0 instrumented:

EgorBo commented Feb 12, 2023 • edited

EgorBo commented Feb 12, 2023

EgorBo commented Feb 13, 2023

jakobbotsch left a comment

Choose a reason for hiding this comment

AndyAyersMS commented Feb 14, 2023

BruceForstall commented Feb 14, 2023

EgorBo commented Oct 23, 2022 •

edited

EgorBo commented Oct 23, 2022 •

edited

EgorBo commented Feb 10, 2023 •

edited

EgorBo commented Feb 12, 2023 •

edited

EgorBo commented Feb 12, 2023 •

edited

EgorBo commented Feb 12, 2023 •

edited