Short-term fix for step into RhpUniversalTransition_DebugStepTailCall

According to recent findings current Windows DIA stackwalker malfunctions at the address ReturnFromUniversalTransition_DebugStepTailCall because it misinterprets the label for the beginning of a new method and assumes that RSP / ESP points to the return address. As the ingestion of an updated DIA version is a longer-term process I'm proposing to temporarily fix this by exporting a temporary variable holding the return address instead.

[tfs-changeset: 1653716]

4 files changed, 58 insertions, 5 deletions

diff --git a/src/Native/Runtime/amd64/CallDescrWorker.S b/src/Native/Runtime/amd64/CallDescrWorker.S
index b9bc36462..6a980b394 100644
--- a/src/Native/Runtime/amd64/CallDescrWorker.S
+++ b/src/Native/Runtime/amd64/CallDescrWorker.S
@@ -6,7 +6,17 @@
 #include <unixasmmacros.inc>
 
 NESTED_ENTRY RhCallDescrWorker, _TEXT, NoHandler
-ALTERNATE_ENTRY ReturnFromCallDescrThunk
+LOCAL_LABEL(ReturnFromCallDescrThunk):
+
     // UNIXTODO: Implement this function
     int 3
 NESTED_END RhCallDescrWorker, _TEXT
+
+    .text
+
+    .align      8
+
+C_FUNC(PointerToReturnFromCallDescrThunk):
+    .quad       LOCAL_LABEL(ReturnFromCallDescrThunk)
+
+    .global     C_FUNC(PointerToReturnFromCallDescrThunk)
diff --git a/src/Native/Runtime/amd64/CallDescrWorker.asm b/src/Native/Runtime/amd64/CallDescrWorker.asm
index 31c996893..a0d07ea45 100644
--- a/src/Native/Runtime/amd64/CallDescrWorker.asm
+++ b/src/Native/Runtime/amd64/CallDescrWorker.asm
@@ -56,8 +56,14 @@ StackCopyLoop:                          ; copy the arguments to stack top-down t
         movdqa  xmm3, [rax + 30h]       ;
 DoCall:
         call    qword ptr [rbx + OFFSETOF__CallDescrData__pTarget]     ; call target function
-LABELED_RETURN_ADDRESS ReturnFromCallDescrThunk ; Symbol used to identify thunk call to managed function so the special 
-                                                ; case unwinder can unwind through this function
+
+ReturnFromCallDescrThunk label proc
+
+        ; Symbol used to identify thunk call to managed function so the special 
+        ; case unwinder can unwind through this function. Sadly we cannot directly
+        ; export this symbol right now because it confuses DIA unwinder to believe
+        ; it's the beginning of a new method, therefore we export the address
+        ; of an auxiliary variable holding the address instead.
 
         ; Save FP return value
 
@@ -97,4 +103,14 @@ ReturnsDouble:
 
         NESTED_END RhCallDescrWorker, _TEXT
 
+        .const
+
+        align   8
+ 
+PointerToReturnFromCallDescrThunk label qword
+
+        dq      offset ReturnFromCallDescrThunk
+
+        public  PointerToReturnFromCallDescrThunk
+
 end
diff --git a/src/Native/Runtime/amd64/UniversalTransition.S b/src/Native/Runtime/amd64/UniversalTransition.S
index cde0ba53b..cf421c630 100644
--- a/src/Native/Runtime/amd64/UniversalTransition.S
+++ b/src/Native/Runtime/amd64/UniversalTransition.S
@@ -124,7 +124,8 @@ NESTED_ENTRY Rhp\FunctionName, _TEXT, NoHandler
         mov  rsi, r11
         lea  rdi, [rsp + DISTANCE_FROM_CHILDSP_TO_RETURN_BLOCK]
         call r10
-ALTERNATE_ENTRY ReturnFrom\FunctionName
+
+LOCAL_LABEL(ReturnFrom\FunctionName):
 
         // restore fp argument registers
         movdqa          xmm0, [rsp + DISTANCE_FROM_CHILDSP_TO_FP_REGS + 0x00]
@@ -151,6 +152,15 @@ ALTERNATE_ENTRY ReturnFrom\FunctionName
 
 NESTED_END Rhp\FunctionName, _TEXT
 
+        .text
+
+        .align          8
+
+C_FUNC(PointerToReturnFrom\FunctionName):
+        .quad           LOCAL_LABEL(ReturnFrom\FunctionName)
+
+        .global         C_FUNC(PointerToReturnFrom\FunctionName)
+
 .endm // UNIVERSAL_TRANSITION
 
         // To enable proper step-in behavior in the debugger, we need to have two instances
diff --git a/src/Native/Runtime/amd64/UniversalTransition.asm b/src/Native/Runtime/amd64/UniversalTransition.asm
index 09f12373a..83aebaf6d 100644
--- a/src/Native/Runtime/amd64/UniversalTransition.asm
+++ b/src/Native/Runtime/amd64/UniversalTransition.asm
@@ -126,7 +126,12 @@ endif ; TRASH_SAVED_ARGUMENT_REGISTERS
         mov  rdx, r11
         lea  rcx, [rsp + DISTANCE_FROM_CHILDSP_TO_RETURN_BLOCK]
         call r10
-LABELED_RETURN_ADDRESS ReturnFrom&FunctionName
+
+ReturnFrom&FunctionName label proc
+
+        ; We cannot make the label public as that tricks DIA stackwalker into thinking
+        ; it's the beginning of a method. For this reason we export the address
+        ; by means of an auxiliary variable.
 
         ; restore fp argument registers
         movdqa          xmm0, [rsp + DISTANCE_FROM_CHILDSP_TO_FP_REGS      ]
@@ -150,6 +155,18 @@ LABELED_RETURN_ADDRESS ReturnFrom&FunctionName
 
 NESTED_END Rhp&FunctionName, _TEXT
 
+        .const
+
+        align           8
+
+PointerToReturnFrom&FunctionName label qword
+
+        dq              offset ReturnFrom&FunctionName
+
+        public          PointerToReturnFrom&FunctionName
+
+        .code
+
         endm
 
         ; To enable proper step-in behavior in the debugger, we need to have two instances