Interface devirt #10192

AndyAyersMS · 2017-03-15T16:39:59Z

Three commits, first has VM changes, second has jit changes, third has tests.

AndyAyersMS · 2017-03-15T16:42:09Z

@jkotas @JosephTremoulet PTAL
cc @dotnet/jit-contrib

More details in the per-commit messages.

For interface calls, when jitting, this invokes the same paths in the runtime as a stub dispatch would, and so may trigger cache updates. It is not clear yet if this update is desirable or should be suppressed. The impact on the caches is similar to what would be without interface call devirtualization, and opportunities for devirtualization are still somewhat scarce, so it may not be easy to find cases where performance differs.

Will post some stats shortly, and workup the desktop equivalent and run tests there too.

AndyAyersMS · 2017-03-15T17:56:44Z

Data from jit-diff on frameworks. A total of 397 interface calls were devirtualized.

Total bytes of diff: 53 (0.00 % of base)
    diff is a regression.

Total byte diff includes 0 bytes from reconciling methods
        Base had    0 unique methods,        0 unique bytes
        Diff had    0 unique methods,        0 unique bytes

Top file regressions by size (bytes):
         520 : System.Linq.Parallel.dasm (0.08 % of base)
         252 : Microsoft.CodeAnalysis.VisualBasic.dasm (0.01 % of base)
         212 : System.Net.Http.dasm (0.10 % of base)
         148 : System.Security.Cryptography.Algorithms.dasm (0.19 % of base)
          98 : System.Security.Cryptography.X509Certificates.dasm (0.13 % of base)

Top file improvements by size (bytes):
       -1297 : System.Private.CoreLib.dasm (-0.04 % of base)
         -76 : System.IO.FileSystem.dasm (-0.12 % of base)
         -60 : Microsoft.Win32.Registry.dasm (-0.27 % of base)
         -31 : System.Collections.dasm (-0.03 % of base)
         -14 : System.Net.Requests.dasm (-0.01 % of base)

19 total files with size differences (7 improved, 12 regressed).

Top method regessions by size (bytes):
          97 : System.Private.CoreLib.dasm - System.Threading.Tasks.Task:WaitAllBlockingCore(ref,int,struct):bool
          61 : Microsoft.CodeAnalysis.CSharp.dasm - Microsoft.CodeAnalysis.CSharp.Emit.PEModuleBuilder:GetExportedTypes(struct):ref:this
          59 : System.Security.Cryptography.Algorithms.dasm - Internal.Cryptography.CngCommon:HashData(ref,int,int,struct):ref
          59 : System.Security.Cryptography.Algorithms.dasm - Internal.Cryptography.CngCommon:HashData(ref,struct):ref
          59 : System.Security.Cryptography.Cng.dasm - Internal.Cryptography.CngCommon:HashData(ref,int,int,struct):ref

Top method improvements by size (bytes):
         -42 : System.Private.CoreLib.dasm - System.Reflection.RuntimeMethodInfo:GetObjectData(ref,struct):this
         -31 : System.Private.CoreLib.dasm - System.Runtime.InteropServices.WindowsRuntime.DictionaryKeyEnumerator`2[__Canon,__Canon][System.__Canon,System.__Canon]:System.Collections.IEnumerator.get_Current():ref:this
         -31 : System.Private.CoreLib.dasm - System.Runtime.InteropServices.WindowsRuntime.DictionaryValueEnumerator`2[__Canon,__Canon][System.__Canon,System.__Canon]:System.Collections.IEnumerator.get_Current():ref:this
         -31 : System.Private.CoreLib.dasm - System.Runtime.InteropServices.WindowsRuntime.EnumeratorToIteratorAdapter`1[__Canon][System.__Canon]:System.Runtime.InteropServices.WindowsRuntime.IBindableIterator.get_Current():ref:this
         -31 : System.Private.CoreLib.dasm - System.Runtime.InteropServices.WindowsRuntime.ReadOnlyDictionaryKeyEnumerator`2[__Canon,__Canon][System.__Canon,System.__Canon]:System.Collections.IEnumerator.get_Current():ref:this

371 total methods with size differences (292 improved, 79 regressed).

Cherry-picked good looking diff where we can now dead code a runtime lookup:

; Assembly listing for method System.Collections.Concurrent.ConcurrentStack`1
;  [__Canon][System.__Canon]:System.Collections.IEnumerable.GetEnumerator():ref:this

; ** Before **
; Total bytes of code 63
G_M45156_IG01:
       push     rsi
       sub      rsp, 48
       mov      qword ptr [rsp+28H], rcx
       mov      rsi, rcx
G_M45156_IG02:
       mov      rcx, qword ptr [rsi]
       mov      rdx, qword ptr [rcx+56]
       mov      rdx, qword ptr [rdx]
       mov      r11, qword ptr [rdx+40]
       test     r11, r11
       jne      SHORT G_M45156_IG03
       lea      rdx, [(reloc)]
       call     CORINFO_HELP_RUNTIMEHANDLE_CLASS
       mov      r11, rax
G_M45156_IG03:
       mov      rcx, rsi
       mov      rax, qword ptr [r11]
       cmp      dword ptr [rcx], ecx
G_M45156_IG04:
       add      rsp, 48
       pop      rsi
       rex.jmp  rax

; ** After **
; Devirtualized interface call to System.Collections.Generic.IEnumerable`1[__Canon]:GetEnumerator; 
;  now direct call to System.Collections.Concurrent.ConcurrentStack`1[__Canon][System.__Canon]:GetEnumerator [final method]
;
; Total bytes of code 38
G_M45158_IG01:
       push     rsi
       sub      rsp, 16
       mov      qword ptr [rsp+08H], rcx
       mov      rsi, rcx
G_M45158_IG02:
       mov      rdx, qword ptr [rsi]
       mov      rdx, gword ptr [rsi+8]
       mov      rcx, rsi
       lea      rax, [(reloc)]
G_M45158_IG03:
       add      rsp, 16
       pop      rsi
       rex.jmp  rax

AndyAyersMS · 2017-03-15T18:26:18Z

Not sure yet what is up with the failing windows pri1 tests. I don't normally run the full pri-1 suite locally. Investigating.

jkotas · 2017-03-15T20:10:24Z

src/vm/jitinterface.cpp

+        //
+        // For generic interface methods we must have an ownerType to
+        // safely devirtualize.
+        if (ownerType != nullptr)


May need o check here that the type implements the interface, similar to what is done for the non-interface case?

Yes, that check is needed. The failing tests have instances where casts are missing and the jit ends up seeing an invalid type.

; Loader\classloader\TypeGeneratorTests\TypeGeneratorTest908\Generated908\Generated908.exe ; Framework:MethodCallingTest() ; local0 is type object IL_0113 06 ldloc.0 IL_0114 25 dup IL_0115 6f 28 00 00 06 callvirt 0x6000028

What is the best way to do this?

Using CanCastToInterface seems to work.

AndyAyersMS · 2017-03-16T15:33:57Z

Updated with additional check on VM side. Full Pri1 suite passes locally.

JosephTremoulet

LGTM, nothing but nits.

JosephTremoulet · 2017-03-16T15:29:17Z

src/vm/jitinterface.cpp

    TypeHandle DerivedClsHnd(derivedClass);
    MethodTable* pDerivedMT = DerivedClsHnd.GetMethodTable();
    _ASSERTE(pDerivedMT->IsRestored() && pDerivedMT->IsFullyLoaded());
-
+        


Nit: inserted trailing whitespace

JosephTremoulet · 2017-03-16T15:31:01Z

src/vm/jitinterface.cpp

+        // exactly derived class. It is up to the jit to determine whether
+        // directly calling this method is correct.
+        pDevirtMD = pDerivedMT->GetMethodDescForSlot(slot);
+        _ASSERTE(pDevirtMD->IsRestored());


Redundant with the assert 3 lines down; did you mean to have both?

Since it is common to both virtual and interface, we only want the second one... will fix.

JosephTremoulet · 2017-03-16T16:03:18Z

tests/src/JIT/opt/Devirtualization/simple.cs

+{
+    new public int F() { return 33; }
+    override public int G() { return 22; }
+


You could consider adding an interface method that B explicitly implements (int Iz.H() { return 11; })

You could also consider adding interfaces that B implements and Z re-implements (for various combinations of implementing explicitly/name-matching-virtual/name-matching-nonvirtual/inheriting-from-B).

JosephTremoulet · 2017-03-16T16:04:31Z

tests/src/JIT/opt/Devirtualization/simple.cs

+    public static int Main()
+    {
+        //       B:F            Z:G            B:F           B:G
+        return Fx(new Z()) + Gy(new Z()) + Fx(new B()) + Gy(new B()) - 65;


Can we get ((Ix)new Z()).F() as well, or is there something special about having the cast to interface happen implicitly for argument passing?

I'll look into this. If C# can elide emitting an actual cast that version should also lead to devirtualization.

We currently lose types in checkcast. I have a partial fix worked up but need to understand (if downcasts are explicit) which casts add type info and which casts lose type info and avoid losing type info.

There are some interface casts that change behaviour, like casting a dictionary to IDictionary changes the enumerator behaviour (from KeyValuePair to DictionaryEntry); may or maynot before something to watch for?

@benaadams I believe that is a C# language feature. At runtime an object only has one type and casting does not alter the object in any way.

Looks like downcasts (to supertype) aren't explicit. Will update the test with some more cases, though I am not sure I will get every possible variation.

AndyAyersMS · 2017-03-16T19:36:38Z

Windows debug failure was a timeout in tailcall_v4_hijacking.

Windows release failures don't repro locally. Only one of the cases does any interface call devirtualization, and it looks to be correct.

Hate to do this but I am going to retry....

@dotnet-bot retest Windows_NT x64 Release Priority 1 Build and Test
@dotnet-bot retest Windows_NT x64 Debug Build and Test

Extend resolveVirtualMethod to take in the owner type needed to resolve shared generic interface calls, and implement VM support for interface call devirtualization. Add a check that the class presented by the jit implements the interface, to catch cases where the IL is missing type casts. Update jit GUID, SPMI and zap to reflect the changed signature.

Unblock VM queries where the base class is an interface (queries where the implementing class is also an interface are still blocked as no resolution is possible). Pass along the context handle that the jit got from getCallInfo.

AndyAyersMS · 2017-03-16T21:15:26Z

Retest was green.

Updates per @JosephTremoulet comments.

AndyAyersMS · 2017-03-17T00:09:22Z

Ported changes over to destktop, initial testing looks good there too.

JosephTremoulet · 2017-03-17T04:49:13Z

tests/src/JIT/opt/Devirtualization/simple.cs

+{
+    new public int F() { return 13; }
+    override public int G() { return 17; }
+    int Iz.H() { return 19; }


If Iz had a couple more methods, B and Z could implement and re-implement them similar to how they do F and G (shadowing and overriding). I don't know/recall what the various C#isms map to in the MSIL for slots and overrides well enough to know if that would be testing anything novel over what you have.

JosephTremoulet · 2017-03-17T04:49:46Z

Updates look good with one comment, still green from me.

AndyAyersMS · 2017-03-17T05:24:21Z

@jkotas any other feedback?

AndyAyersMS · 2017-03-17T05:29:45Z

@dotnet-bot test Windows_NT x64 corefx_baseline

jkotas · 2017-03-17T05:29:46Z

LGTM

AndyAyersMS · 2017-03-17T06:05:46Z

@RussKeldorph what is the trigger phrase for corefx testing?

AndyAyersMS · 2017-03-17T06:08:25Z

Ah, looks like what I used above worked, but the test leg seems broken....

RussKeldorph · 2017-03-17T16:19:43Z

@AndyAyersMS Yes, corefx jobs are broken due to 14207f4 breaking change. Apparently we have to wait until corefx ingests a coreclr with those bits before the break can be resolved. @rahku is tracking it.

dnfclas added the cla-already-signed label Mar 15, 2017

AndyAyersMS added area-CodeGen optimization labels Mar 15, 2017

jkotas reviewed Mar 15, 2017

View reviewed changes

AndyAyersMS force-pushed the InterfaceDevirt branch from d5d5d01 to 913795e Compare March 16, 2017 15:33

JosephTremoulet approved these changes Mar 16, 2017

View reviewed changes

AndyAyersMS added 3 commits March 16, 2017 13:51

Interface call devirtualization: jit side of changes

2c8b3ae

Unblock VM queries where the base class is an interface (queries where the implementing class is also an interface are still blocked as no resolution is possible). Pass along the context handle that the jit got from getCallInfo.

Add some test cases

4e02195

AndyAyersMS force-pushed the InterfaceDevirt branch from 913795e to 4e02195 Compare March 16, 2017 21:14

JosephTremoulet reviewed Mar 17, 2017

View reviewed changes

AndyAyersMS merged commit 8208cb5 into dotnet:master Mar 17, 2017

AndyAyersMS deleted the InterfaceDevirt branch March 17, 2017 18:04

karelz modified the milestone: 2.0.0 Aug 28, 2017

AndyAyersMS mentioned this pull request Jan 31, 2020

JIT: devirtualization next steps dotnet/runtime#7541

Open

26 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Interface devirt #10192

Interface devirt #10192

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

jkotas Mar 15, 2017

AndyAyersMS Mar 15, 2017

AndyAyersMS Mar 15, 2017

jkotas Mar 16, 2017

AndyAyersMS commented Mar 16, 2017

JosephTremoulet left a comment

JosephTremoulet Mar 16, 2017

JosephTremoulet Mar 16, 2017

AndyAyersMS Mar 16, 2017

JosephTremoulet Mar 16, 2017

JosephTremoulet Mar 16, 2017

AndyAyersMS Mar 16, 2017

benaadams Mar 16, 2017 •

edited

Loading

AndyAyersMS Mar 16, 2017

AndyAyersMS Mar 16, 2017

AndyAyersMS commented Mar 16, 2017 •

edited

Loading

AndyAyersMS commented Mar 16, 2017

AndyAyersMS commented Mar 17, 2017

JosephTremoulet Mar 17, 2017

JosephTremoulet commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

jkotas commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

RussKeldorph commented Mar 17, 2017

Interface devirt #10192

Interface devirt #10192

Conversation

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

AndyAyersMS commented Mar 15, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AndyAyersMS commented Mar 16, 2017

JosephTremoulet left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

benaadams Mar 16, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AndyAyersMS commented Mar 16, 2017 • edited Loading

AndyAyersMS commented Mar 16, 2017

AndyAyersMS commented Mar 17, 2017

Choose a reason for hiding this comment

JosephTremoulet commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

jkotas commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

AndyAyersMS commented Mar 17, 2017

RussKeldorph commented Mar 17, 2017

benaadams Mar 16, 2017 •

edited

Loading

AndyAyersMS commented Mar 16, 2017 •

edited

Loading