Revert VSD stubs changes

Extensive benchmarking has shown that the new model for VSD stubs is
causing steady state performance degradation for code using a lot of
virtual calls. It wasn't showing up in the plaintext benchmark I have
used as the driver of the change.
This change looses part of the startup perf improvements, but I am
planning to add back the Lookup stubs in a follow up change. Those
should not be perf critical and they are likely the ones where the
startup improvements are gained from.

Author: janvorli

src/coreclr/vm/amd64/asmconstants.h

@@ -590,39 +590,6 @@ ASMCONSTANTS_C_ASSERT(CallCountingStubData__TargetForMethod == offsetof(CallCoun
 #define CallCountingStubData__TargetForThresholdReached 0x10
 ASMCONSTANTS_C_ASSERT(CallCountingStubData__TargetForThresholdReached == offsetof(CallCountingStubData, TargetForThresholdReached))
 
-#define               LookupStubData__DispatchToken 0x00
-ASMCONSTANTS_C_ASSERT(LookupStubData__DispatchToken == offsetof(LookupStubData, DispatchToken))
-
-#define               LookupStubData__ResolveWorkerTarget 0x08
-ASMCONSTANTS_C_ASSERT(LookupStubData__ResolveWorkerTarget == offsetof(LookupStubData, ResolveWorkerTarget))
-
-#define               DispatchStubData__ExpectedMT 0x00
-ASMCONSTANTS_C_ASSERT(DispatchStubData__ExpectedMT == offsetof(DispatchStubData, ExpectedMT))
-
-#define               DispatchStubData__ImplTarget 0x08
-ASMCONSTANTS_C_ASSERT(DispatchStubData__ImplTarget == offsetof(DispatchStubData, ImplTarget))
-
-#define               DispatchStubData__FailTarget 0x10
-ASMCONSTANTS_C_ASSERT(DispatchStubData__FailTarget == offsetof(DispatchStubData, FailTarget))
-
-#define               ResolveStubData__HashedToken 0x08
-ASMCONSTANTS_C_ASSERT(ResolveStubData__HashedToken == offsetof(ResolveStubData, HashedToken))
-
-#define               ResolveStubData__CacheAddress 0x00
-ASMCONSTANTS_C_ASSERT(ResolveStubData__CacheAddress == offsetof(ResolveStubData, CacheAddress))
-
-#define               ResolveStubData__Token 0x10
-ASMCONSTANTS_C_ASSERT(ResolveStubData__Token == offsetof(ResolveStubData, Token))
-
-#define               ResolveStubData__Counter 0x0c
-ASMCONSTANTS_C_ASSERT(ResolveStubData__Counter == offsetof(ResolveStubData, Counter))
-
-#define               ResolveStubData__ResolveWorkerTarget 0x18
-ASMCONSTANTS_C_ASSERT(ResolveStubData__ResolveWorkerTarget == offsetof(ResolveStubData, ResolveWorkerTarget))
-
-#define CALL_STUB_CACHE_MASK_ASM 0xfff
-ASMCONSTANTS_C_ASSERT(CALL_STUB_CACHE_MASK_ASM == CALL_STUB_CACHE_MASK)
-
 #undef ASMCONSTANTS_RUNTIME_ASSERT
 #undef ASMCONSTANTS_C_ASSERT
 #ifndef UNIX_AMD64_ABI

src/coreclr/vm/amd64/thunktemplates.S

@@ -29,50 +29,3 @@ LEAF_ENTRY CallCountingStubCode, _TEXT
     LOCAL_LABEL(CountReachedZero):
         jmp    QWORD PTR [rip + DATA_SLOT(CallCountingStub, TargetForThresholdReached)]
 LEAF_END_MARKED CallCountingStubCode, _TEXT
-
-LEAF_ENTRY LookupStubCode, _TEXT
-       push   QWORD PTR [rip + DATA_SLOT(LookupStub, DispatchToken)]
-       jmp    QWORD PTR [rip + DATA_SLOT(LookupStub, ResolveWorkerTarget)]
-LEAF_END_MARKED LookupStubCode, _TEXT
-
-LEAF_ENTRY DispatchStubCode, _TEXT
-       mov    rax,QWORD PTR [rip + DATA_SLOT(DispatchStub, ExpectedMT)]
-PATCH_LABEL DispatchStubCode_ThisDeref
-       cmp    QWORD PTR [rdi],rax;
-       jne    LOCAL_LABEL(Fail)
-       jmp    QWORD PTR [rip + DATA_SLOT(DispatchStub, ImplTarget)]
-    LOCAL_LABEL(Fail):
-       jmp    QWORD PTR [rip + DATA_SLOT(DispatchStub, FailTarget)]
-LEAF_END_MARKED DispatchStubCode, _TEXT
-
-LEAF_ENTRY ResolveStubCode, _TEXT
-PATCH_LABEL ResolveStubCode_ResolveEntry
-    LOCAL_LABEL(Resolve):
-        push   rdx
-        mov    r10,QWORD PTR [rip + DATA_SLOT(ResolveStub, CacheAddress)]
-PATCH_LABEL ResolveStubCode_ThisDeref
-        mov    rax,QWORD PTR [rdi]
-        mov    rdx,rax
-        shr    rax,12
-        add    rax,rdx
-        xor    eax,DWORD PTR [rip + DATA_SLOT(ResolveStub, HashedToken)]
-        and    eax, CALL_STUB_CACHE_MASK_ASM * 8
-        mov    rax,QWORD PTR [r10+rax*1]
-        mov    r10,QWORD PTR [rip + DATA_SLOT(ResolveStub, Token)]
-        cmp    rdx,QWORD PTR [rax]
-        jne    LOCAL_LABEL(Miss)
-        cmp    r10,QWORD PTR [rax+8]
-        jne    LOCAL_LABEL(Miss)
-        pop    rdx
-        jmp    QWORD PTR [rax+0x10]
-PATCH_LABEL ResolveStubCode_FailEntry
-        add    DWORD PTR [rip + DATA_SLOT(ResolveStub, Counter)], -1
-        jge    LOCAL_LABEL(Resolve)
-        or     r11, 1 // SDF_ResolveBackPatch
-PATCH_LABEL ResolveStubCode_SlowEntry
-        push   rdx
-        mov    r10,QWORD PTR [rip + DATA_SLOT(ResolveStub, Token)]
-    LOCAL_LABEL(Miss):
-        push   rax
-        jmp    QWORD PTR [rip + DATA_SLOT(ResolveStub, ResolveWorkerTarget)]
-LEAF_END_MARKED ResolveStubCode, _TEXT

src/coreclr/vm/amd64/thunktemplates.asm

@@ -31,50 +31,4 @@ LEAF_ENTRY CallCountingStubCode, _TEXT
         jmp    QWORD PTR [DATA_SLOT(CallCountingStub, TargetForThresholdReached)]
 LEAF_END_MARKED CallCountingStubCode, _TEXT
 
-LEAF_ENTRY LookupStubCode, _TEXT
-       push   QWORD PTR [DATA_SLOT(LookupStub, DispatchToken)]
-       jmp    QWORD PTR [DATA_SLOT(LookupStub, ResolveWorkerTarget)]
-LEAF_END_MARKED LookupStubCode, _TEXT
-
-LEAF_ENTRY DispatchStubCode, _TEXT
-       mov    rax,QWORD PTR [DATA_SLOT(DispatchStub, ExpectedMT)]
-PATCH_LABEL DispatchStubCode_ThisDeref
-       cmp    QWORD PTR [rcx],rax;
-       jne    Fail
-       jmp    QWORD PTR [DATA_SLOT(DispatchStub, ImplTarget)]
-    Fail:
-       jmp    QWORD PTR [DATA_SLOT(DispatchStub, FailTarget)]
-LEAF_END_MARKED DispatchStubCode, _TEXT
-
-LEAF_ENTRY ResolveStubCode, _TEXT
-PATCH_LABEL ResolveStubCode_ResolveEntry
-        push   rdx
-        mov    r10,QWORD PTR [DATA_SLOT(ResolveStub, CacheAddress)]
-PATCH_LABEL ResolveStubCode_ThisDeref
-        mov    rax,QWORD PTR [rcx]
-        mov    rdx,rax
-        shr    rax,12
-        add    rax,rdx
-        xor    eax,DWORD PTR [DATA_SLOT(ResolveStub, HashedToken)]
-        and    eax, CALL_STUB_CACHE_MASK_ASM * 8
-        mov    rax,QWORD PTR [r10+rax*1]
-        mov    r10,QWORD PTR [DATA_SLOT(ResolveStub, Token)]
-        cmp    rdx,QWORD PTR [rax]
-        jne    Miss
-        cmp    r10,QWORD PTR [rax+8]
-        jne    Miss
-        pop    rdx
-        jmp    QWORD PTR [rax+10h]
-PATCH_LABEL ResolveStubCode_FailEntry
-        add    DWORD PTR [DATA_SLOT(ResolveStub, Counter)], -1
-        jge    ResolveStubCode
-        or     r11, 1; SDF_ResolveBackPatch
-PATCH_LABEL ResolveStubCode_SlowEntry
-        push   rdx
-        mov    r10,QWORD PTR [DATA_SLOT(ResolveStub, Token)]
-Miss:
-        push   rax
-        jmp    QWORD PTR [DATA_SLOT(ResolveStub, ResolveWorkerTarget)]
-LEAF_END_MARKED ResolveStubCode, _TEXT
-
         end