Implement FixupPrecode for Arm64

Author: ramarag

src/inc/jithelpers.h

@@ -315,7 +315,7 @@
 
     JITHELPER(CORINFO_HELP_EE_PRESTUB,          ThePreStub,                 CORINFO_HELP_SIG_NO_ALIGN_STUB)
 
-#if defined(_TARGET_X86_) || defined(_TARGET_AMD64_) || defined(_TARGET_ARM_)
+#if defined(HAS_FIXUP_PRECODE) 
     JITHELPER(CORINFO_HELP_EE_PRECODE_FIXUP,    PrecodeFixupThunk,          CORINFO_HELP_SIG_NO_ALIGN_STUB)
 #else
     JITHELPER(CORINFO_HELP_EE_PRECODE_FIXUP,    NULL,                       CORINFO_HELP_SIG_NO_ALIGN_STUB)

src/vm/arm64/asmconstants.h

@@ -147,6 +147,19 @@ ASMCONSTANTS_C_ASSERT(CONTEXT_Pc == offsetof(T_CONTEXT,Pc))
 ASMCONSTANTS_C_ASSERT(SIZEOF__FaultingExceptionFrame        == sizeof(FaultingExceptionFrame));
 ASMCONSTANTS_C_ASSERT(FaultingExceptionFrame__m_fFilterExecuted == offsetof(FaultingExceptionFrame, m_fFilterExecuted));
 
+#define SIZEOF__FixupPrecode                 24
+#define Offset_PrecodeChunkIndex             15
+#define Offset_MethodDescChunkIndex          14
+#define MethodDesc_ALIGNMENT_SHIFT           3
+#define FixupPrecode_ALIGNMENT_SHIFT_1       3
+#define FixupPrecode_ALIGNMENT_SHIFT_2       4
+
+ASMCONSTANTS_C_ASSERT(SIZEOF__FixupPrecode == sizeof(FixupPrecode));
+ASMCONSTANTS_C_ASSERT(Offset_PrecodeChunkIndex == offsetof(FixupPrecode, m_PrecodeChunkIndex));
+ASMCONSTANTS_C_ASSERT(Offset_MethodDescChunkIndex == offsetof(FixupPrecode, m_MethodDescChunkIndex));
+ASMCONSTANTS_C_ASSERT(MethodDesc_ALIGNMENT_SHIFT == MethodDesc::ALIGNMENT_SHIFT);
+ASMCONSTANTS_C_ASSERT((1<<FixupPrecode_ALIGNMENT_SHIFT_1) + (1<<FixupPrecode_ALIGNMENT_SHIFT_2)  == sizeof(FixupPrecode));
+
 #ifndef CROSSGEN_COMPILE
 #define ResolveCacheElem__target      0x10
 #define ResolveCacheElem__pNext       0x18

src/vm/arm64/asmhelpers.asm

@@ -193,9 +193,25 @@ Done
         NESTED_END
 
 ; ------------------------------------------------------------------
-; ARM64TODO: Implement PrecodeFixupThunk when PreCode is Enabled
+; The call in fixup precode initally points to this function.
+; The pupose of this function is to load the MethodDesc and forward the call to prestub.
         NESTED_ENTRY PrecodeFixupThunk
-        brk        #0
+
+        ; x12 = FixupPrecode *
+        ; On Exit
+        ; x12 = MethodDesc*
+        ; x13, x14 Trashed
+        ; Inline computation done by FixupPrecode::GetMethodDesc()
+        ldrb    w13, [x12, #Offset_PrecodeChunkIndex]              ; m_PrecodeChunkIndex
+        ldrb    w14, [x12, #Offset_MethodDescChunkIndex]           ; m_MethodDescChunkIndex
+
+        add     x12,x12,w13,uxtw #FixupPrecode_ALIGNMENT_SHIFT_1
+        add     x13,x12,w13,uxtw #FixupPrecode_ALIGNMENT_SHIFT_2
+        ldr     x13, [x13,#SIZEOF__FixupPrecode]
+        add     x12,x13,w14,uxtw #MethodDesc_ALIGNMENT_SHIFT
+
+        b ThePreStub
+
         NESTED_END
 ; ------------------------------------------------------------------
 

src/vm/arm64/cgencpu.h

@@ -38,8 +38,8 @@ class ComCallMethodDesc;
 
 #define USE_INDIRECT_CODEHEADER
 
-//#define HAS_FIXUP_PRECODE                       1
-//#define HAS_FIXUP_PRECODE_CHUNKS                1
+#define HAS_FIXUP_PRECODE                       1
+#define HAS_FIXUP_PRECODE_CHUNKS                1
 
 // ThisPtrRetBufPrecode one is necessary for closed delegates over static methods with return buffer
 #define HAS_THISPTR_RETBUF_PRECODE              1
@@ -613,44 +613,76 @@ typedef DPTR(NDirectImportPrecode) PTR_NDirectImportPrecode;
 
 struct FixupPrecode {
 
-    static const int Type = 0xfc;
+    static const int Type = 0x0C;
 
-    // mov r12, pc
-    // ldr pc, [pc, #4]     ; =m_pTarget
+    // adr x12, #0 
+    // ldr x11, [pc, #12]     ; =m_pTarget
+    // br  x11
     // dcb m_MethodDescChunkIndex
     // dcb m_PrecodeChunkIndex
+    // 2 byte padding
     // dcd m_pTarget
-    WORD    m_rgCode[3];
+
+
+    UINT32  m_rgCode[3];
+    BYTE    padding[2];
     BYTE    m_MethodDescChunkIndex;
     BYTE    m_PrecodeChunkIndex;
     TADDR   m_pTarget;
 
     void Init(MethodDesc* pMD, LoaderAllocator *pLoaderAllocator, int iMethodDescChunkIndex = 0, int iPrecodeChunkIndex = 0);
+    void InitCommon()
+    {
+        WRAPPER_NO_CONTRACT;
+        int n = 0;
+
+        m_rgCode[n++] = 0x1000000C;   // adr x12, #0
+        m_rgCode[n++] = 0x5800006B;   // ldr x11, [pc, #12]     ; =m_pTarget
+
+        _ASSERTE((UINT32*)&m_pTarget == &m_rgCode[n + 2]);
+
+        m_rgCode[n++] = 0xD61F0160;   // br  x11
+
+        _ASSERTE(n == _countof(m_rgCode));
+    }
 
     TADDR GetBase()
     {
-        _ASSERTE(!"ARM64:NYI");
-        return NULL;
+        LIMITED_METHOD_CONTRACT;
+        SUPPORTS_DAC;
+
+        return dac_cast<TADDR>(this) + (m_PrecodeChunkIndex + 1) * sizeof(FixupPrecode);
     }
 
     TADDR GetMethodDesc();
 
     PCODE GetTarget()
     {
-        _ASSERTE(!"ARM64:NYI");
-        return NULL;
+        LIMITED_METHOD_DAC_CONTRACT;
+        return m_pTarget;
     }
 
     BOOL SetTargetInterlocked(TADDR target, TADDR expected)
     {
-        _ASSERTE(!"ARM64:NYI");
-        return NULL;
+        CONTRACTL
+        {
+            THROWS;
+            GC_TRIGGERS;
+        }
+        CONTRACTL_END;
+
+        EnsureWritableExecutablePages(&m_pTarget);
+        return (TADDR)InterlockedCompareExchange64(
+            (LONGLONG*)&m_pTarget, (TADDR)target, (TADDR)expected) == expected;
     }
 
     static BOOL IsFixupPrecodeByASM(PCODE addr)
     {
-        _ASSERTE(!"ARM64:NYI");
-        return NULL;
+        PTR_DWORD pInstr = dac_cast<PTR_DWORD>(PCODEToPINSTR(addr));
+        return
+            (pInstr[0] == 0x1000000C) &&
+            (pInstr[1] == 0x5800006B) &&
+            (pInstr[2] == 0xD61F0160);
     }
 
 #ifdef FEATURE_PREJIT

src/vm/arm64/stubs.cpp

@@ -494,14 +494,22 @@ void HelperMethodFrame::UpdateRegDisplay(const PREGDISPLAY pRD)
 
 TADDR FixupPrecode::GetMethodDesc()
 {
-    _ASSERTE(!"ARM64:NYI");
-    return NULL;
+    LIMITED_METHOD_DAC_CONTRACT;
+
+    // This lookup is also manually inlined in PrecodeFixupThunk assembly code
+    TADDR base = *PTR_TADDR(GetBase());
+    if (base == NULL)
+        return NULL;
+    return base + (m_MethodDescChunkIndex * MethodDesc::ALIGNMENT);
 }
 
 #ifdef DACCESS_COMPILE
 void FixupPrecode::EnumMemoryRegions(CLRDataEnumMemoryFlags flags)
 {
-    _ASSERTE(!"ARM64:NYI");
+	SUPPORTS_DAC;
+	DacEnumMemoryRegion(dac_cast<TADDR>(this), sizeof(FixupPrecode));
+
+	DacEnumMemoryRegion(GetBase(), sizeof(TADDR));
 }
 #endif // DACCESS_COMPILE
 
@@ -574,19 +582,78 @@ void NDirectImportPrecode::Fixup(DataImage *image)
 
 void FixupPrecode::Init(MethodDesc* pMD, LoaderAllocator *pLoaderAllocator, int iMethodDescChunkIndex /*=0*/, int iPrecodeChunkIndex /*=0*/)
 {
-    _ASSERTE(!"ARM64:NYI");
+    WRAPPER_NO_CONTRACT;
+    
+    InitCommon();
+
+    // Initialize chunk indices only if they are not initialized yet. This is necessary to make MethodDesc::Reset work.
+    if (m_PrecodeChunkIndex == 0)
+    {
+        _ASSERTE(FitsInU1(iPrecodeChunkIndex));
+        m_PrecodeChunkIndex = static_cast<BYTE>(iPrecodeChunkIndex);
+    }
+
+    if (iMethodDescChunkIndex != -1)
+    {
+        if (m_MethodDescChunkIndex == 0)
+        {
+            _ASSERTE(FitsInU1(iMethodDescChunkIndex));
+            m_MethodDescChunkIndex = static_cast<BYTE>(iMethodDescChunkIndex);
+        }
+
+        if (*(void**)GetBase() == NULL)
+            *(void**)GetBase() = (BYTE*)pMD - (iMethodDescChunkIndex * MethodDesc::ALIGNMENT);
+    }
+
+    _ASSERTE(GetMethodDesc() == (TADDR)pMD);
+
+    if (pLoaderAllocator != NULL)
+    {
+        m_pTarget = GetEEFuncEntryPoint(PrecodeFixupThunk);
+    }
 }
 
 #ifdef FEATURE_NATIVE_IMAGE_GENERATION
 // Partial initialization. Used to save regrouped chunks.
 void FixupPrecode::InitForSave(int iPrecodeChunkIndex)
 {
-    _ASSERTE(!"ARM64:NYI");
+    STANDARD_VM_CONTRACT;
+
+    InitCommon();
+
+    _ASSERTE(FitsInU1(iPrecodeChunkIndex));
+    m_PrecodeChunkIndex = static_cast<BYTE>(iPrecodeChunkIndex);
+    // The rest is initialized in code:FixupPrecode::Fixup
 }
 
 void FixupPrecode::Fixup(DataImage *image, MethodDesc * pMD)
 {
-    _ASSERTE(!"ARM64:NYI");
+    STANDARD_VM_CONTRACT;
+
+    // Note that GetMethodDesc() does not return the correct value because of 
+    // regrouping of MethodDescs into hot and cold blocks. That's why the caller
+    // has to supply the actual MethodDesc
+
+    SSIZE_T mdChunkOffset;
+    ZapNode * pMDChunkNode = image->GetNodeForStructure(pMD, &mdChunkOffset);
+    ZapNode * pHelperThunk = image->GetHelperThunk(CORINFO_HELP_EE_PRECODE_FIXUP);
+
+    image->FixupFieldToNode(this, offsetof(FixupPrecode, m_pTarget), pHelperThunk);
+
+    // Set the actual chunk index
+    FixupPrecode * pNewPrecode = (FixupPrecode *)image->GetImagePointer(this);
+
+    size_t mdOffset = mdChunkOffset - sizeof(MethodDescChunk);
+    size_t chunkIndex = mdOffset / MethodDesc::ALIGNMENT;
+    _ASSERTE(FitsInU1(chunkIndex));
+    pNewPrecode->m_MethodDescChunkIndex = (BYTE)chunkIndex;
+
+    // Fixup the base of MethodDescChunk
+    if (m_PrecodeChunkIndex == 0)
+    {
+        image->FixupFieldToNode(this, (BYTE *)GetBase() - (BYTE *)this,
+            pMDChunkNode, sizeof(MethodDescChunk));
+    }
 }
 #endif // FEATURE_NATIVE_IMAGE_GENERATION
 
@@ -618,7 +685,20 @@ BOOL DoesSlotCallPrestub(PCODE pCode)
 {
     PTR_DWORD pInstr = dac_cast<PTR_DWORD>(PCODEToPINSTR(pCode));
 
-    // ARM64TODO: Check for FixupPrecode
+    //FixupPrecode
+#if defined(HAS_FIXUP_PRECODE)
+    if (FixupPrecode::IsFixupPrecodeByASM(pCode))
+    {
+        PCODE pTarget = dac_cast<PTR_FixupPrecode>(pInstr)->m_pTarget;
+        
+        if (isJump(pTarget))
+        {
+            pTarget = decodeJump(pTarget);
+        }
+
+        return pTarget == (TADDR)PrecodeFixupThunk;
+    }
+#endif
 
     // StubPrecode
     if (pInstr[0] == 0x10000089 && // adr x9, #16
@@ -627,7 +707,10 @@ BOOL DoesSlotCallPrestub(PCODE pCode)
     {
         PCODE pTarget = dac_cast<PTR_StubPrecode>(pInstr)->m_pTarget;
 
-        // ARM64TODO: implement for NGen case
+        if (isJump(pTarget))
+        {
+            pTarget = decodeJump(pTarget);
+        }
 
         return pTarget == GetPreStubEntryPoint();
     }

src/vm/precode.cpp

@@ -433,7 +433,7 @@ BOOL Precode::SetTargetInterlocked(PCODE target)
     // SetTargetInterlocked does not modify code on ARM so the flush instruction cache is
     // not necessary.
     //
-#if !defined(_TARGET_ARM_)
+#if !defined(_TARGET_ARM_) && !defined(_TARGET_ARM64_)
     if (ret) {
         FlushInstructionCache(GetCurrentProcess(),this,SizeOf());
     }