ARM64: Enable Function Fragment

Fixes https://github.com/dotnet/coreclr/issues/6064
Fixes https://github.com/dotnet/coreclr/issues/6122

Basically this enable function fragment.
Unwind data for arm64 can express only 1MB range.
For the large function, we should split unwind data for each fragment of
function. Other than the first fragment (the host region), each fragment
has a phantom prolog starting with "end_c(0xe5)".
I confirmed that this implementation aligns with window's unwinder, and
tested with COMPlus_JitSplitFunctionSize=32 for a few test cases which
split both main and funclets.

Author: kyulee1

src/jit/unwind.h

@@ -30,6 +30,7 @@ const unsigned MAX_EPILOG_SIZE_BYTES = 40;
 const unsigned MAX_PROLOG_SIZE_BYTES = 100;
 const unsigned MAX_EPILOG_SIZE_BYTES = 100;
 #define UWC_END 0xE4            // "end" unwind code
+#define UWC_END_C 0xE5          // "end_c" unwind code
 #define UW_MAX_FRAGMENT_SIZE_BYTES          (1U << 20)
 #define UW_MAX_CODE_WORDS_COUNT             31
 #define UW_MAX_EPILOG_START_INDEX           0x3FFU

src/jit/unwindarm.cpp

@@ -996,9 +996,11 @@ void            UnwindFragmentInfo::AddEpilog()
 void            UnwindFragmentInfo::CopyPrologCodes(UnwindFragmentInfo* pCopyFrom)
 {
     ufiPrologCodes.CopyFrom(&pCopyFrom->ufiPrologCodes);
+#ifdef _TARGET_ARM64_
+    ufiPrologCodes.AddCode(UWC_END_C);
+#endif
 }
 
-
 // Split the epilog codes that currently exist in 'pSplitFrom'. The ones that represent
 // epilogs that start at or after the location represented by 'emitLoc' are removed
 // from 'pSplitFrom' and moved to this fragment. Note that this fragment should not have

src/vm/codeman.cpp

@@ -858,39 +858,34 @@ BOOL IsFunctionFragment(TADDR baseAddress, PTR_RUNTIME_FUNCTION pFunctionEntry)
     // 1. Prolog only: The host record. Epilog Count and E bit are all 0.
     // 2. Prolog and some epilogs: The host record with acompannying epilog-only records
     // 3. Epilogs only: First unwind code is Phantom prolog (Starting with an end_c, indicating an empty prolog)
-    // 4. No prologs or epilogs: Epilog Count = 1 and Epilog Start Index points end_c. (as if it's case #2 with empty epilog codes) 
+    // 4. No prologs or epilogs: First unwind code is Phantom prolog  (Starting with an end_c, indicating an empty prolog)
     //
 
     int EpilogCount = (int)(unwindHeader >> 22) & 0x1F;
     int CodeWords = unwindHeader >> 27;
     PTR_DWORD pUnwindCodes = (PTR_DWORD)(baseAddress + pFunctionEntry->UnwindData);
+    // Skip header.
+    pUnwindCodes++;
+
+    // Skip extended header.
     if ((CodeWords == 0) && (EpilogCount == 0))
-        pUnwindCodes++;
-    BOOL Ebit = (unwindHeader >> 21) & 0x1;
-    if (Ebit)
     {
-        // EpilogCount is the index of the first unwind code that describes the one and only epilog
-        // The unwind codes immediatelly follow the unwindHeader
+        EpilogCount = (*pUnwindCodes) & 0xFFFF;
         pUnwindCodes++;
     }
-    else if (EpilogCount != 0)
+
+    // Skip epilog scopes.
+    BOOL Ebit = (unwindHeader >> 21) & 0x1;
+    if (!Ebit && (EpilogCount != 0))
     {
         // EpilogCount is the number of exception scopes defined right after the unwindHeader
-        pUnwindCodes += EpilogCount+1;
+        pUnwindCodes += EpilogCount;
     }
-    else
-    {
-        return FALSE;
-    }
-
-    if ((*pUnwindCodes & 0xFF) == 0xE5) // Phantom prolog
-        return TRUE;
-        
 
+    return ((*pUnwindCodes & 0xFF) == 0xE5);
 #else
     PORTABILITY_ASSERT("IsFunctionFragnent - NYI on this platform");
 #endif
-    return FALSE;
 }
 
 #endif // EXCEPTION_DATA_SUPPORTS_FUNCTION_FRAGMENTS

tests/arm64/Tests.lst

@@ -36572,35 +36572,35 @@ RelativePath=JIT\jit64\opt\cse\HugeArray1\HugeArray1.cmd
 WorkingDir=JIT\jit64\opt\cse\HugeArray1
 Expected=0
 MaxAllowedDurationSeconds=600
-Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING;GCSTRESS_FAIL;ISSUE_6064
+Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING
 HostStyle=0
 [hugeexpr1.cmd_5297]
 RelativePath=JIT\jit64\opt\cse\hugeexpr1\hugeexpr1.cmd
 WorkingDir=JIT\jit64\opt\cse\hugeexpr1
 Expected=0
 MaxAllowedDurationSeconds=600
-Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING;GCSTRESS_FAIL;ISSUE_6064
+Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING
 HostStyle=0
 [HugeField1.cmd_5298]
 RelativePath=JIT\jit64\opt\cse\HugeField1\HugeField1.cmd
 WorkingDir=JIT\jit64\opt\cse\HugeField1
 Expected=0
 MaxAllowedDurationSeconds=600
-Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING;GCSTRESS_FAIL;ISSUE_6064
+Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING
 HostStyle=0
 [HugeField2.cmd_5299]
 RelativePath=JIT\jit64\opt\cse\HugeField2\HugeField2.cmd
 WorkingDir=JIT\jit64\opt\cse\HugeField2
 Expected=0
 MaxAllowedDurationSeconds=600
-Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING;GCSTRESS_FAIL;ISSUE_6122
+Categories=Pri0;JIT;EXPECTED_PASS;LONG_RUNNING
 HostStyle=0
 [hugeSimpleExpr1.cmd_5300]
 RelativePath=JIT\jit64\opt\cse\hugeSimpleExpr1\hugeSimpleExpr1.cmd
 WorkingDir=JIT\jit64\opt\cse\hugeSimpleExpr1
 Expected=0
 MaxAllowedDurationSeconds=600
-Categories=Pri0;JIT;EXPECTED_PASS;GCSTRESS_FAIL;ISSUE_6064
+Categories=Pri0;JIT;EXPECTED_PASS
 HostStyle=0
 [mixedexpr1_d_loop_try.cmd_5301]
 RelativePath=JIT\jit64\opt\cse\mixedexpr1_d_loop_try\mixedexpr1_d_loop_try.cmd