Enable inlining P/Invokes into try blocks with no catch or filter clauses

src/coreclr/inc/corinfo.h

@@ -3242,4 +3242,9 @@ class ICorDynamicInfo : public ICorStaticInfo
 //
 #define IMAGE_REL_BASED_REL_THUMB_MOV32_PCREL   0x14
 
+/**********************************************************************************/
+#ifdef TARGET_64BIT
+#define USE_PER_FRAME_PINVOKE_INIT
+#endif
+
 #endif // _COR_INFO_H_

src/coreclr/jit/importer.cpp

@@ -8429,41 +8429,37 @@ bool Compiler::impCanPInvokeInlineCallSite(BasicBlock* block)
         return true;
     }
 
-#ifdef TARGET_64BIT
-    // On 64-bit platforms, we disable pinvoke inlining inside of try regions.
-    // Note that this could be needed on other architectures too, but we
-    // haven't done enough investigation to know for sure at this point.
-    //
-    // Here is the comment from JIT64 explaining why:
-    //   [VSWhidbey: 611015] - because the jitted code links in the
-    //   Frame (instead of the stub) we rely on the Frame not being
-    //   'active' until inside the stub.  This normally happens by the
-    //   stub setting the return address pointer in the Frame object
-    //   inside the stub.  On a normal return, the return address
-    //   pointer is zeroed out so the Frame can be safely re-used, but
-    //   if an exception occurs, nobody zeros out the return address
-    //   pointer.  Thus if we re-used the Frame object, it would go
-    //   'active' as soon as we link it into the Frame chain.
-    //
-    //   Technically we only need to disable PInvoke inlining if we're
-    //   in a handler or if we're in a try body with a catch or
-    //   filter/except where other non-handler code in this method
-    //   might run and try to re-use the dirty Frame object.
-    //
-    //   A desktop test case where this seems to matter is
-    //   jit\jit64\ebvts\mcpp\sources2\ijw\__clrcall\vector_ctor_dtor.02\deldtor_clr.exe
+#ifdef USE_PER_FRAME_PINVOKE_INIT
+    // For platforms that use per-P/Invoke InlinedCallFrame initialization,
+    // we can't inline P/Invokes inside of try blocks where we can resume execution in the same function.
+    // The runtime can correctly unwind out of an InlinedCallFrame and out of managed code. However,
+    // it cannot correctly unwind out of an InlinedCallFrame and stop at that frame without also unwinding
+    // at least one managed frame. In particular, the runtime struggles to restore non-volatile registers
+    // from the top-most unmanaged call before the InlinedCallFrame. As a result, the runtime does not support
+    // re-entering the same method frame as the InlinedCallFrame after an exception in unmanaged code.
     if (block->hasTryIndex())
     {
         // This does not apply to the raw pinvoke call that is inside the pinvoke
         // ILStub. In this case, we have to inline the raw pinvoke call into the stub,
         // otherwise we would end up with a stub that recursively calls itself, and end
         // up with a stack overflow.
+        // This works correctly because the runtime never emits a catch block in a managed-to-native
+        // IL stub. If the runtime ever emits a catch block into a managed-to-native stub when using
+        // P/Invoke helpers, this condition will need to be revisited.
         if (opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB) && opts.ShouldUsePInvokeHelpers())
         {
             return true;
         }
 
-        return false;
+        for (BasicBlock* ehsucc : block->GetEHSuccs(this))
+        {
+            if (ehsucc->hasHndIndex() && ehsucc->bbCatchTyp != BBCT_FAULT && ehsucc->bbCatchTyp != BBCT_FINALLY)
+            {
+                return false;
+            }
+        }
+
+        return true;
     }
 #endif // TARGET_64BIT
 

src/coreclr/jit/lower.cpp

@@ -4281,6 +4281,7 @@ GenTree* Lowering::CreateFrameLinkUpdate(FrameLinkAction action)
 // Return Value:
 //    none
 //
+// See the usages for USE_PER_FRAME_PINVOKE_INIT for more information.
 void Lowering::InsertPInvokeMethodProlog()
 {
     noway_assert(comp->info.compUnmanagedCallCountWithGCTransition);
@@ -4377,13 +4378,16 @@ void Lowering::InsertPInvokeMethodProlog()
     // --------------------------------------------------------
     // On 32-bit targets, CORINFO_HELP_INIT_PINVOKE_FRAME initializes the PInvoke frame and then pushes it onto
     // the current thread's Frame stack. On 64-bit targets, it only initializes the PInvoke frame.
+    // As a result, don't push the frame onto the frame stack here for any 64-bit targets
     CLANG_FORMAT_COMMENT_ANCHOR;
 
 #ifdef TARGET_64BIT
+#ifdef USE_PER_FRAME_PINVOKE_INIT
+    // For IL stubs, we push the frame once even when we're doing per-pinvoke init.
     if (comp->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB))
+#endif // USE_PER_FRAME_PINVOKE_INIT
     {
-        // Push a frame - if we are NOT in an IL stub, this is done right before the call
-        // The init routine sets InlinedCallFrame's m_pNext, so we just set the thead's top-of-stack
+        // Push a frame. The init routine sets InlinedCallFrame's m_pNext, so we just set the thead's top-of-stack
         GenTree* frameUpd = CreateFrameLinkUpdate(PushFrame);
         firstBlockRange.InsertBefore(insertionPoint, LIR::SeqTree(comp, frameUpd));
         ContainCheckStoreIndir(frameUpd->AsStoreInd());
@@ -4443,9 +4447,10 @@ void Lowering::InsertPInvokeMethodEpilog(BasicBlock* returnBB DEBUGARG(GenTree*
     // this in the epilog for IL stubs; for non-IL stubs the frame is popped after every PInvoke call.
     CLANG_FORMAT_COMMENT_ANCHOR;
 
-#ifdef TARGET_64BIT
+#ifdef USE_PER_FRAME_PINVOKE_INIT
+    // For IL stubs, we push the frame once even when we're doing per-pinvoke init
     if (comp->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB))
-#endif // TARGET_64BIT
+#endif // USE_PER_FRAME_PINVOKE_INIT
     {
         GenTree* frameUpd = CreateFrameLinkUpdate(PopFrame);
         returnBlockRange.InsertBefore(insertionPoint, LIR::SeqTree(comp, frameUpd));
@@ -4601,7 +4606,7 @@ void Lowering::InsertPInvokeCallProlog(GenTreeCall* call)
     // contains PInvokes; on 64-bit targets this is necessary in non-stubs.
     CLANG_FORMAT_COMMENT_ANCHOR;
 
-#ifdef TARGET_64BIT
+#ifdef USE_PER_FRAME_PINVOKE_INIT
     if (!comp->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB))
     {
         // Set the TCB's frame to be the one we just created.
@@ -4613,7 +4618,7 @@ void Lowering::InsertPInvokeCallProlog(GenTreeCall* call)
         BlockRange().InsertBefore(insertBefore, LIR::SeqTree(comp, frameUpd));
         ContainCheckStoreIndir(frameUpd->AsStoreInd());
     }
-#endif // TARGET_64BIT
+#endif // USE_PER_FRAME_PINVOKE_INIT
 
     // IMPORTANT **** This instruction must be the last real instruction ****
     // It changes the thread's state to Preemptive mode
@@ -4679,7 +4684,7 @@ void Lowering::InsertPInvokeCallEpilog(GenTreeCall* call)
     // happens after every PInvoke call in non-stubs. 32-bit targets instead mark the frame as inactive.
     CLANG_FORMAT_COMMENT_ANCHOR;
 
-#ifdef TARGET_64BIT
+#ifdef USE_PER_FRAME_PINVOKE_INIT
     if (!comp->opts.jitFlags->IsSet(JitFlags::JIT_FLAG_IL_STUB))
     {
         tree = CreateFrameLinkUpdate(PopFrame);
@@ -4703,7 +4708,7 @@ void Lowering::InsertPInvokeCallEpilog(GenTreeCall* call)
 
     BlockRange().InsertBefore(insertionPoint, constantZero, storeCallSiteTracker);
     ContainCheckStoreLoc(storeCallSiteTracker);
-#endif // TARGET_64BIT
+#endif // USE_PER_FRAME_PINVOKE_INIT
 }
 
 //------------------------------------------------------------------------

src/coreclr/vm/exceptionhandling.cpp

@@ -1776,8 +1776,10 @@ CLRUnwindStatus ExceptionTracker::ProcessOSExceptionNotification(
             // InlinedCallFrames (ICF) are allocated, initialized and linked to the Frame chain
             // by the code generated by the JIT for a method containing a PInvoke.
             //
-            // JIT generates code that links in the ICF at the start of the method and unlinks it towards
-            // the method end. Thus, ICF is present on the Frame chain at any given point so long as the
+            // On platforms where USE_PER_FRAME_PINVOKE_INIT is not defined,
+            // the JIT generates code that links in the ICF
+            // at the start of the method and unlinks it towards the method end.
+            // Thus, ICF is present on the Frame chain at any given point so long as the
             // method containing the PInvoke is on the stack.
             //
             // Now, if the method containing ICF catches an exception, we will reset the Frame chain
@@ -1815,13 +1817,16 @@ CLRUnwindStatus ExceptionTracker::ProcessOSExceptionNotification(
             //    below the callerSP for which we will invoke ExceptionUnwind.
             //
             //    Thus, ICF::ExceptionUnwind should not do anything significant. If any of these assumptions
-            //    break, then the next best thing will be to make the JIT link/unlink the frame dynamically.
+            //    break, then the next best thing will be to make the JIT link/unlink the frame dynamically
             //
-            // If the current method executing is from precompiled ReadyToRun code, then the above is no longer
-            // applicable because each PInvoke is wrapped by calls to the JIT_PInvokeBegin and JIT_PInvokeEnd
-            // helpers, which push and pop the ICF to the current thread. Unlike jitted code, the ICF is not
-            // linked during the method prolog, and unlinked at the epilog (it looks more like the X64 case).
+            // If the current method executing is from precompiled ReadyToRun code, each PInvoke is wrapped
+            // by calls to the JIT_PInvokeBegin and JIT_PInvokeEnd helpers,
+            // which push and pop the ICF to the current thread. The ICF is not
+            // linked during the method prolog, and unlinked at the epilog.
             // In that case, we need to unlink the ICF during unwinding here.
+            // On platforms where USE_PER_FRAME_PINVOKE_INIT is defined, the JIT generates code that links in
+            // the ICF immediately before and after a PInvoke in non-IL-stubs, like ReadyToRun.
+            // See the usages for USE_PER_FRAME_PINVOKE_INIT for more information.
 
             if (fTargetUnwind && (pFrame->GetVTablePtr() == InlinedCallFrame::GetMethodFrameVPtr()))
             {
@@ -1840,9 +1845,18 @@ CLRUnwindStatus ExceptionTracker::ProcessOSExceptionNotification(
                     // to the JIT_PInvokeBegin and JIT_PInvokeEnd helpers, which push and pop the ICF on the thread. The
                     // ICF is not linked at the method prolog and unlined at the epilog when running R2R code. Since the
                     // JIT_PInvokeEnd helper will be skipped, we need to unlink the ICF here. If the executing method
-                    // has another pinovoke, it will re-link the ICF again when the JIT_PInvokeBegin helper is called
+                    // has another pinvoke, it will re-link the ICF again when the JIT_PInvokeBegin helper is called.
 
-                    if (ExecutionManager::IsReadyToRunCode(((InlinedCallFrame*)pFrame)->m_pCallerReturnAddress))
+                    TADDR returnAddress = ((InlinedCallFrame*)pFrame)->m_pCallerReturnAddress;
+#ifdef USE_PER_FRAME_PINVOKE_INIT
+                    // If we're setting up the frame for each P/Invoke for the given platform,
+                    // then we do this for all P/Invokes except ones in IL stubs.
+                    if (!ExecutionManager::GetCodeMethodDesc(returnAddress)->IsILStub())
+#else
+                    // If we aren't setting up the frame for each P/Invoke (instead setting up once per method),
+                    // then ReadyToRun code is the only code using the per-P/Invoke logic.
+                    if (ExecutionManager::IsReadyToRunCode(returnAddress))
+#endif
                     {
                         pICFForUnwindTarget = pICFForUnwindTarget->Next();
                     }

src/coreclr/vm/i386/excepx86.cpp

@@ -2982,6 +2982,8 @@ void ResumeAtJitEH(CrawlFrame* pCf,
         // Check that the InlinedCallFrame is in the method with the exception handler. There can be other
         // InlinedCallFrame somewhere up the call chain that is not related to the current exception
         // handling.
+
+        // See the usages for USE_PER_FRAME_PINVOKE_INIT for more information.
 
 #ifdef DEBUG
         TADDR handlerFrameSP = pCf->GetRegisterSet()->SP;
@@ -2994,10 +2996,22 @@ void ResumeAtJitEH(CrawlFrame* pCf,
                                                                      NULL /* StackwalkCacheUnwindInfo* */);
         _ASSERTE(unwindSuccess);
 
-        if (((TADDR)pThread->m_pFrame < pCf->GetRegisterSet()->SP) && ExecutionManager::IsReadyToRunCode(((InlinedCallFrame*)pThread->m_pFrame)->m_pCallerReturnAddress))
+        if (((TADDR)pThread->m_pFrame < pCf->GetRegisterSet()->SP))
         {
-            _ASSERTE((TADDR)pThread->m_pFrame >= handlerFrameSP);
-            pThread->m_pFrame->Pop(pThread);
+            TADDR returnAddress = ((InlinedCallFrame*)pThread->m_pFrame)->m_pCallerReturnAddressl
+#ifdef USE_PER_FRAME_PINVOKE_INIT
+            // If we're setting up the frame for each P/Invoke for the given platform,
+            // then we do this for all P/Invokes except ones in IL stubs.
+            if (!ExecutionManager::GetCodeMethodDesc(returnAddress)->IsILStub())
+#else
+            // If we aren't setting up the frame for each P/Invoke (instead setting up once per method),
+            // then ReadyToRun code is the only code using the per-P/Invoke logic.
+            if (ExecutionManager::IsReadyToRunCode(returnAddress))
+#endif
+            {
+                _ASSERTE((TADDR)pThread->m_pFrame >= handlerFrameSP);
+                pThread->m_pFrame->Pop(pThread);
+            }
         }
     }
 

src/tests/baseservices/exceptions/exceptioninterop/ExceptionInterop.cs

@@ -122,4 +122,38 @@ public static void ThrowNativeExceptionAndCatchInFrameWithFinally()
 
         Assert.True(caughtException);
     }
+
+    [Fact]
+    [PlatformSpecific(TestPlatforms.Windows)]
+    [SkipOnMono("Exception interop not supported on Mono.")]
+    public static void ThrowNativeExceptionInFrameWithFinallyCatchInOuterFrame()
+    {
+        bool caughtException = false;
+        try
+        {
+            ThrowInFrameWithFinally();
+        }
+        catch
+        {
+            caughtException = true;
+        }
+
+        Assert.True(caughtException);
+
+        [MethodImpl(MethodImplOptions.NoInlining)]
+        static void ThrowInFrameWithFinally()
+        {
+            try
+            {
+                ThrowException();
+            }
+            finally
+            {
+                // Try calling another P/Invoke in the finally block before the catch
+                // to make sure we have everything set up
+                // to recover from the exceptional control flow.
+                NativeFunction();
+            }
+        }
+    }
 }