MRT StackFrameIterator improvements

These changes are targeted at completing/hardening the runtime's
support for the new UniversalTransition and CallDescr thunks and also
clarifying the invariants that hold throughout stack walker operation.

[tfs-changeset: 1568546]

3 files changed, 26 insertions, 46 deletions

diff --git a/src/Native/Runtime/i386/AsmMacros.inc b/src/Native/Runtime/i386/AsmMacros.inc
index 8a249d2c1..ffda6561e 100644
--- a/src/Native/Runtime/i386/AsmMacros.inc
+++ b/src/Native/Runtime/i386/AsmMacros.inc
@@ -85,31 +85,21 @@ endm
 ;;
 ;; EAX is trashed by this macro.
 ;;
-;; If <NoModeSwitch> is set, threadReg is ignored. As a result, <NoModeSwitch> can only be used by helpers
-;; which guarantee that they will never transition into pre-emptive mode and will never examine the m_pThread
-;; field of the frame.
-;;
-PUSH_COOP_PINVOKE_FRAME macro threadReg, NoModeSwitch
+PUSH_COOP_PINVOKE_FRAME macro threadReg
     lea         eax, [ebp + 8]                      ; get the ESP of the caller
     push        eax                                 ; save ESP
     push        edi
     push        esi
     push        ebx
     push        PTFF_SAVE_ALL_PRESERVED + PTFF_SAVE_RSP
-ifb <NoModeSwitch>
     push        threadReg                           ; Thread * (unused by stackwalker)
-else
-    push        0
-endif
     mov         eax, [ebp + 0]                      ; Find previous EBP value
     push        eax                                 ; save EBP
     mov         eax, [ebp + 4]                      ; Find the return address
     push        eax                                 ; save m_RIP
 
     lea         eax, [esp + 0]                      ; EAX == address of frame
-ifb <NoModeSwitch>
     mov         [threadReg + OFFSETOF__Thread__m_pHackPInvokeTunnel], eax ; Link frame to thread
-endif
 endm
 
 ;;
diff --git a/src/Native/Runtime/i386/AsmOffsetsCpu.h b/src/Native/Runtime/i386/AsmOffsetsCpu.h
index c214c06fb..be2d04abc 100644
--- a/src/Native/Runtime/i386/AsmOffsetsCpu.h
+++ b/src/Native/Runtime/i386/AsmOffsetsCpu.h
@@ -9,7 +9,7 @@
 //
 // NOTE: the offsets MUST be in hex notation WITHOUT the 0x prefix
 
-PLAT_ASM_SIZEOF(b4, ExInfo)
+PLAT_ASM_SIZEOF(b8, ExInfo)
 PLAT_ASM_OFFSET(0, ExInfo, m_pPrevExInfo)
 PLAT_ASM_OFFSET(4, ExInfo, m_pExContext)
 PLAT_ASM_OFFSET(8, ExInfo, m_exception)
@@ -17,7 +17,7 @@ PLAT_ASM_OFFSET(0c, ExInfo, m_kind)
 PLAT_ASM_OFFSET(0d, ExInfo, m_passNumber)
 PLAT_ASM_OFFSET(10, ExInfo, m_idxCurClause)
 PLAT_ASM_OFFSET(14, ExInfo, m_frameIter)
-PLAT_ASM_OFFSET(b0, ExInfo, m_notifyDebuggerSP)
+PLAT_ASM_OFFSET(b4, ExInfo, m_notifyDebuggerSP)
 
 PLAT_ASM_OFFSET(0, PInvokeTransitionFrame, m_RIP)
 PLAT_ASM_OFFSET(4, PInvokeTransitionFrame, m_FramePointer)
@@ -25,7 +25,7 @@ PLAT_ASM_OFFSET(8, PInvokeTransitionFrame, m_pThread)
 PLAT_ASM_OFFSET(0c, PInvokeTransitionFrame, m_dwFlags)
 PLAT_ASM_OFFSET(10, PInvokeTransitionFrame, m_PreservedRegs)
 
-PLAT_ASM_SIZEOF(9c, StackFrameIterator)
+PLAT_ASM_SIZEOF(a0, StackFrameIterator)
 PLAT_ASM_OFFSET(08, StackFrameIterator, m_FramePointer)
 PLAT_ASM_OFFSET(0c, StackFrameIterator, m_ControlPC)
 PLAT_ASM_OFFSET(10, StackFrameIterator, m_RegDisplay)
diff --git a/src/Native/Runtime/i386/UniversalTransition.asm b/src/Native/Runtime/i386/UniversalTransition.asm
index 7abd1fc63..1f7a13634 100644
--- a/src/Native/Runtime/i386/UniversalTransition.asm
+++ b/src/Native/Runtime/i386/UniversalTransition.asm
@@ -34,24 +34,26 @@ ifdef FEATURE_DYNAMIC_CODE
 ;;
 
 ;
-; Stack frame layout (from lower addresses to higher addresses):
+; Frame layout is:
 ;
-; [callee return]
-; [pinvoke frame, 20h]
-; [in edx (argument register from the caller)]
-; [in ecx (argument register from the caller)]
-; [ConservativelyReportedScratchSpace 8h]
-;   -- On input (i.e., when control jumps to RhpUniversalTransition), the low 4 bytes of
-;      this area contain the address of the callee and the high 4 bytes of this area contain
-;      the extra argument to pass to the callee.
-; [ptr to pinvoke frame 4h]
-; [saved ebp register]
-; [caller return addr]
-; [stack-passed arguments from the caller]
+;   {StackPassedArgs}                           ChildSP+018     CallerSP+000
+;   {CallerRetaddr}                             ChildSP+014     CallerSP-004
+;   {CallerEBP}                                 ChildSP+010     CallerSP-008
+;   {ReturnBlock (0x8 bytes)}                   ChildSP+008     CallerSP-010
+;    -- On input (i.e., when control jumps to RhpUniversalTransition), the low 4 bytes of
+;       the ReturnBlock area holds the address of the callee and the high 4 bytes holds the
+;       extra argument to pass to the callee.
+;   {IntArgRegs (edx,ecx) (0x8 bytes)}          ChildSP+000     CallerSP-018
+;   {CalleeRetaddr}                             ChildSP-004     CallerSP-01c
 ;
-; Note: The callee receives a pointer to the pushed edx value, and the callee has
-; knowledge of the exact layout of all pieces of the frame that lie at or above the pushed
-; edx value.
+; NOTE: If the frame layout ever changes, the C++ UniversalTransitionStackFrame structure
+; must be updated as well.
+;
+; NOTE: The callee receives a pointer to the base of the pushed IntArgRegs, and the callee
+; has knowledge of the exact layout of the entire frame.
+;
+; NOTE: The stack walker guarantees that conservative GC reporting will be applied to
+; everything between the base of the IntArgRegs and the top of the StackPassedArgs.
 ;
 
 FASTCALL_FUNC RhpUniversalTransition_FAKE_ENTRY, 0        
@@ -60,34 +62,22 @@ FASTCALL_FUNC RhpUniversalTransition_FAKE_ENTRY, 0
         mov         ebp, esp
         push eax
         push eax
-        push eax
 ALTERNATE_ENTRY RhpUniversalTransition@0
         push ecx
         push edx
 
-        ; Build the frame that the stack walker will use to unwind through this function.  The
-        ; <NoModeSwitch> flag indicates that this function never uses Enable/DisablePreemptiveGC,
-        ; implying that frame address does not need to be recorded in the current thread object.
-        PUSH_COOP_PINVOKE_FRAME notUsed, <NoModeSwitch>
-
-        ;; Stash the pinvoke frame's address immediately on top of the old ebp value. This
-        ;; position is important; the stack frame iterator knows about this setup.
-.erre MANAGED_CALLOUT_THUNK_TRANSITION_FRAME_POINTER_OFFSET eq -4
-        mov [ebp-4], esp
-
         ;
         ; Call out to the target, while storing and reporting arguments to the GC.
         ;
-        mov  eax, [ebp-0Ch]  ; Get the address of the callee 
-        mov  edx, [ebp-8]    ; Get the extra argument to pass to the callee
-        lea  ecx, [ebp-14h]  ; Get pointer to edx value pushed above
+        mov  eax, [ebp-8]    ; Get the address of the callee 
+        mov  edx, [ebp-4]    ; Get the extra argument to pass to the callee
+        lea  ecx, [ebp-10h]  ; Get pointer to edx value pushed above
         call eax
 LABELED_RETURN_ADDRESS ReturnFromUniversalTransition
 
-        POP_COOP_PINVOKE_FRAME
         pop edx
         pop ecx
-        add esp, 12
+        add esp, 8
         pop ebp
         jmp eax