[CodeGen][ObjC] Try/Catch support for WebAssembly

[HendrikHuebner]

Related to #169043 and part of a broader effort to support targetting WebAssembly for ObjectiveC.

This PR is work in progress and adds basic support for generating funclet-style exception handling (try/catch) for WebAssembly. For now, I've set the gxx_wasm personality function. I'm not entirely sure whether we need to define out own.

For these examples, the generated IR matches the IR generated from an equivalent C++ program (Aside from the exception type ID).

extern void might_throw(void);

void test_simple_try_catch(void) {
  @try {
    might_throw();
  } @catch (id e) {
  } @catch (id e) {
  } 
}

void test_catch_specific_type(void) {
  @try {
    might_throw();
  } @catch (id e) {
  }
}

void test_nested_try_catch(void) {
  @try {
    @try {
      might_throw();
    } @catch (id e) {
    }
  } @catch (id e) {
  }
}

[HendrikHuebner]

cc @hmelder

[HendrikHuebner]

TODO: This is taken from CGException.cpp, we may want to refactor it into a shared helper.

[github-actions]

⚠️ C/C++ code formatter, clang-format found issues in your code. ⚠️

You can test this locally with the following command:

git-clang-format --diff origin/main HEAD --extensions h,cpp -- clang/lib/CodeGen/CGCleanup.h clang/lib/CodeGen/CGException.cpp clang/lib/CodeGen/CGObjCRuntime.cpp clang/lib/CodeGen/CodeGenFunction.h clang/lib/Driver/ToolChains/Clang.cpp --diff_from_common_commit

⚠️
The reproduction instructions above might return results for more than one PR
in a stack if you are using a stacked PR workflow. You can limit the results by
changing origin/main to the base branch/commit you want to compare against.
⚠️

View the diff from clang-format here.

diff --git a/clang/lib/CodeGen/CGObjCRuntime.cpp b/clang/lib/CodeGen/CGObjCRuntime.cpp
index aa4622695..4d3bf1c1e 100644
--- a/clang/lib/CodeGen/CGObjCRuntime.cpp
+++ b/clang/lib/CodeGen/CGObjCRuntime.cpp
@@ -157,8 +157,8 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
   CodeGenFunction::FinallyInfo FinallyInfo;
   if (!useFunclets)
     if (const ObjCAtFinallyStmt *Finally = S.getFinallyStmt())
-      FinallyInfo.enter(CGF, Finally->getFinallyBody(),
-                        beginCatchFn, endCatchFn, exceptionRethrowFn);
+      FinallyInfo.enter(CGF, Finally->getFinallyBody(), beginCatchFn,
+                        endCatchFn, exceptionRethrowFn);
   SmallVector<CatchHandler, 8> Handlers;
 
 
@@ -193,9 +193,9 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
 
   if (useFunclets) {
     if (const ObjCAtFinallyStmt *Finally = S.getFinallyStmt()) {
-        if (hasWasmExceptions) {
-          CGF.ErrorUnsupported(Finally, "@finally for WASM");
-        }
+      if (hasWasmExceptions) {
+        CGF.ErrorUnsupported(Finally, "@finally for WASM");
+      }
 
         CodeGenFunction HelperCGF(CGM, /*suppressNewContext=*/true);
         if (!CGF.CurSEHParent)
@@ -217,12 +217,11 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
     }
   }
 
-
   // Emit the try body.
   CGF.EmitStmt(S.getTryBody());
 
   // Leave the try.
-  llvm::BasicBlock* dispatchBlock{};
+  llvm::BasicBlock *dispatchBlock{};
   if (S.getNumCatchStmts())
     dispatchBlock = CGF.popCatchScope();
 
@@ -234,14 +233,15 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
   // each catch handler.
   SaveAndRestore RestoreCurrentFuncletPad(CGF.CurrentFuncletPad);
   llvm::BasicBlock *WasmCatchStartBlock = nullptr;
-  llvm::CatchPadInst* CatchPadInst{};
+  llvm::CatchPadInst *CatchPadInst{};
   if (!!dispatchBlock && hasWasmExceptions) {
     auto *CatchSwitch =
         cast<llvm::CatchSwitchInst>(dispatchBlock->getFirstNonPHIIt());
     WasmCatchStartBlock = CatchSwitch->hasUnwindDest()
                               ? CatchSwitch->getSuccessor(1)
                               : CatchSwitch->getSuccessor(0);
-    CatchPadInst = cast<llvm::CatchPadInst>(WasmCatchStartBlock->getFirstNonPHIIt());
+    CatchPadInst =
+        cast<llvm::CatchPadInst>(WasmCatchStartBlock->getFirstNonPHIIt());
     CGF.CurrentFuncletPad = CatchPadInst;
   }
 
@@ -258,9 +258,11 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
       llvm::BasicBlock::iterator CPICandidate =
           Handler.Block->getFirstNonPHIIt();
       if (CPICandidate != Handler.Block->end()) {
-        if ((CatchPadInst = dyn_cast_or_null<llvm::CatchPadInst>(CPICandidate))) {
+        if ((CatchPadInst =
+                 dyn_cast_or_null<llvm::CatchPadInst>(CPICandidate))) {
           CGF.CurrentFuncletPad = CatchPadInst;
-          CatchPadInst->setOperand(2, CGF.getExceptionSlot().emitRawPointer(CGF));
+          CatchPadInst->setOperand(2,
+                                   CGF.getExceptionSlot().emitRawPointer(CGF));
         }
       }
     }
@@ -318,7 +320,7 @@ void CGObjCRuntime::EmitTryCatchStmt(CodeGenFunction &CGF,
     assert(RethrowBlock != WasmCatchStartBlock && RethrowBlock->empty());
     CGF.Builder.SetInsertPoint(RethrowBlock);
     llvm::Function *RethrowInCatchFn =
-      CGM.getIntrinsic(llvm::Intrinsic::wasm_rethrow);
+        CGM.getIntrinsic(llvm::Intrinsic::wasm_rethrow);
     CGF.EmitNoreturnRuntimeCallOrInvoke(RethrowInCatchFn, {});
   }
 
diff --git a/clang/lib/CodeGen/CodeGenFunction.h b/clang/lib/CodeGen/CodeGenFunction.h
index 86c02f0ac..b284f58bc 100644
--- a/clang/lib/CodeGen/CodeGenFunction.h
+++ b/clang/lib/CodeGen/CodeGenFunction.h
@@ -1307,7 +1307,7 @@ public:
   /// popCatchScope - Pops the catch scope at the top of the EHScope
   /// stack, emitting any required code (other than the catch handlers
   /// themselves).
-  llvm::BasicBlock* popCatchScope();
+  llvm::BasicBlock *popCatchScope();
 
   llvm::BasicBlock *getEHResumeBlock(bool isCleanup);
   llvm::BasicBlock *getEHDispatchBlock(EHScopeStack::stable_iterator scope);

[hmelder]

Thank you!

cc @davidchisnall

we likely need to add a new one to handle ObjectiveC's @finally semantics.

This is currently also broken on Windows (MSVC ABI) so we might be able to fix this one as well.

For now, I've set the gxx_wasm personality function

I'll test your patch tomorrow and adapt the personality function in libobjc2 if needed.

[hmelder]

Given that we emit the same IR instructions for EH on Windows.

[HendrikHuebner]

Thank you!

cc @davidchisnall

we likely need to add a new one to handle ObjectiveC's @finally semantics.

This is currently also broken on Windows (MSVC ABI) so we might be able to fix this one as well.

For now, I've set the gxx_wasm personality function

I'll test your patch tomorrow and adapt the personality function in libobjc2 if needed.

I'm not sure if the personality function is going to be a concern. The WebAssembly VM should handle the unwinding process (see here). However, other runtime function may need to be adapted in libobjc2, as they were for C++.

[triplef]

Just for context: the issue with @finally on Windows is tracked in #43828 and #51899.

[hmelder]

extern void might_throw(void);

void test_simple_try_catch(void) {
  @try {
    might_throw();
  } @catch (id e) {
  } @catch (id e) {
  } 
}

I compiled your example with -target wasm32-unknown-emscripten -fwasm-exceptions -mllvm -wasm-use-legacy-eh=false -fobjc-runtime=gnustep-2.2 -S -emit-llvm and the generated IR looks good.

The generated IR

; Function Attrs: noinline optnone
define hidden void @test_simple_try_catch() #0 personality ptr @__gxx_wasm_personality_v0 {
entry:
  %exn.slot = alloca ptr, align 4
  %e = alloca ptr, align 4
  %e5 = alloca ptr, align 4
  invoke void @might_throw()
          to label %invoke.cont unwind label %catch.dispatch

catch.dispatch:                                   ; preds = %entry
  %0 = catchswitch within none [label %catch.start] unwind to caller

catch.start:                                      ; preds = %catch.dispatch
  %1 = catchpad within %0 [ptr @0, ptr @0]
  %2 = call ptr @llvm.wasm.get.exception(token %1)
  store ptr %2, ptr %exn.slot, align 4
  %3 = call i32 @llvm.wasm.get.ehselector(token %1)
  %4 = call i32 @llvm.eh.typeid.for.p0(ptr @0) #5
  %matches = icmp eq i32 %3, %4
  br i1 %matches, label %catch, label %catch.fallthrough

catch.fallthrough:                                ; preds = %catch.start
  %5 = call i32 @llvm.eh.typeid.for.p0(ptr @0) #5
  %matches1 = icmp eq i32 %3, %5
  br i1 %matches1, label %catch2, label %rethrow

rethrow:                                          ; preds = %catch.fallthrough
  call void @llvm.wasm.rethrow() #4 [ "funclet"(token %1) ]
  unreachable

invoke.cont:                                      ; preds = %entry
  br label %eh.cont

eh.cont:                                          ; preds = %invoke.cont, %catchret.dest6, %catchret.dest
  ret void

catch:                                            ; preds = %catch.start
  %exn = load ptr, ptr %exn.slot, align 4
  %exn.adjusted = call ptr @objc_begin_catch(ptr %exn) #5 [ "funclet"(token %1) ]
  store ptr %exn.adjusted, ptr %e, align 4
  call void @objc_end_catch() #5 [ "funclet"(token %1) ]
  catchret from %1 to label %catchret.dest

catchret.dest:                                    ; preds = %catch
  br label %eh.cont

catch2:                                           ; preds = %catch.fallthrough
  %exn3 = load ptr, ptr %exn.slot, align 4
  %exn.adjusted4 = call ptr @objc_begin_catch(ptr %exn3) #5 [ "funclet"(token %1) ]
  store ptr %exn.adjusted4, ptr %e5, align 4
  call void @objc_end_catch() #5 [ "funclet"(token %1) ]
  catchret from %1 to label %catchret.dest6

catchret.dest6:                                   ; preds = %catch2
  br label %eh.cont
}

[hmelder]

I'm not sure if the personality function is going to be a concern.

AFAIK it is. On WASM, there is only one Unwind phase instead of a search and install phase.

[hmelder]

I'm currently hacking in support for the ObjC personality function (crimes are being commited) and it looks like __gxx_wasm_personality_v0 is hardcoded in the backend. For example WasmEHPrepare.cpp:244.