Skip to content

release/22.x: [CoroSplit] Never collect allocas used by catchpad into frame#194531

Merged
c-rhodes merged 1 commit into
llvm:release/22.xfrom
NewSigma:issue186728
May 26, 2026
Merged

release/22.x: [CoroSplit] Never collect allocas used by catchpad into frame#194531
c-rhodes merged 1 commit into
llvm:release/22.xfrom
NewSigma:issue186728

Conversation

@NewSigma

Copy link
Copy Markdown
Contributor

Windows EH requires exception objects allocated on stack. But there is
no reliable way to identify them. CoroSplit employs a best-effort
algorithm to determine whether allocas persist on the stack or the
frame, which may result in miscompilation when Windows exceptions are
used.
This patch proposes that we treat allocas used by catchpad as exception
objects and never place them on the frame. A verifier check is added to
enforce that operands of catchpad are either constants or allocas.

Close #143235 Close #153949 Close #182584

@llvmbot

llvmbot commented Apr 28, 2026

Copy link
Copy Markdown
Member

@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-coroutines

Author: Weibo He (NewSigma)

Changes

Windows EH requires exception objects allocated on stack. But there is
no reliable way to identify them. CoroSplit employs a best-effort
algorithm to determine whether allocas persist on the stack or the
frame, which may result in miscompilation when Windows exceptions are
used.
This patch proposes that we treat allocas used by catchpad as exception
objects and never place them on the frame. A verifier check is added to
enforce that operands of catchpad are either constants or allocas.

Close #143235 Close #153949 Close #182584


Full diff: https://github.com/llvm/llvm-project/pull/194531.diff

7 Files Affected:

  • (modified) llvm/docs/Coroutines.rst (+3)
  • (modified) llvm/docs/LangRef.rst (+3-2)
  • (modified) llvm/lib/FuzzMutate/RandomIRBuilder.cpp (+5)
  • (modified) llvm/lib/IR/Verifier.cpp (+7)
  • (modified) llvm/lib/Transforms/Coroutines/SpillUtils.cpp (+7)
  • (added) llvm/test/Transforms/Coroutines/coro-alloca-10.ll (+66)
  • (modified) llvm/unittests/FuzzMutate/StrategiesTest.cpp (+2-1)
diff --git a/llvm/docs/Coroutines.rst b/llvm/docs/Coroutines.rst
index 0e6b49c84acee..9fa403e68e13b 100644
--- a/llvm/docs/Coroutines.rst
+++ b/llvm/docs/Coroutines.rst
@@ -2252,4 +2252,7 @@ Areas Requiring Attention
 #. Make required changes to make sure that coroutine optimizations work with
    LTO.
 
+#. In Windows EH, exception objects must be allocated on the stack (see :ref:wineh for details).
+   We identify an exception object as an alloca that has `catchpad` users.
+
 #. More tests, more tests, more tests
diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
index be752b42ea412..1e0293f67e833 100644
--- a/llvm/docs/LangRef.rst
+++ b/llvm/docs/LangRef.rst
@@ -13825,8 +13825,9 @@ ensures that each ``catchpad`` has exactly one predecessor block, and it always
 terminates in a ``catchswitch``.
 
 The ``args`` correspond to whatever information the personality routine
-requires to determine if this is an appropriate handler for the exception. Control
-will transfer to the ``catchpad`` if this is the first appropriate handler for
+requires to determine if this is an appropriate handler for the exception.
+Each operand must be an alloca or a constant.
+Control will transfer to the ``catchpad`` if this is the first appropriate handler for
 the exception.
 
 The ``resultval`` has the type :ref:`token <t_token>` and is used to match the
diff --git a/llvm/lib/FuzzMutate/RandomIRBuilder.cpp b/llvm/lib/FuzzMutate/RandomIRBuilder.cpp
index 8b034137e4702..87392c2e69299 100644
--- a/llvm/lib/FuzzMutate/RandomIRBuilder.cpp
+++ b/llvm/lib/FuzzMutate/RandomIRBuilder.cpp
@@ -332,6 +332,11 @@ static bool isCompatibleReplacement(const Instruction *I, const Use &Operand,
       return false;
     return !Callee->hasParamAttribute(OperandNo, Attribute::ImmArg);
   }
+  case Instruction::CatchPad:
+    // Argument operand must be alloca or constant
+    if (!isa<Constant>(Replacement) && !isa<AllocaInst>(Replacement))
+      return false;
+    break;
   default:
     break;
   }
diff --git a/llvm/lib/IR/Verifier.cpp b/llvm/lib/IR/Verifier.cpp
index bb552861130d2..bdca6800eddfe 100644
--- a/llvm/lib/IR/Verifier.cpp
+++ b/llvm/lib/IR/Verifier.cpp
@@ -4911,6 +4911,13 @@ void Verifier::visitCatchPadInst(CatchPadInst &CPI) {
   Check(&*BB->getFirstNonPHIIt() == &CPI,
         "CatchPadInst not the first non-PHI instruction in the block.", &CPI);
 
+  Check(llvm::all_of(CPI.arg_operands(),
+                     [](Use &U) {
+                       auto *V = U.get();
+                       return isa<Constant>(V) || isa<AllocaInst>(V);
+                     }),
+        "Argument operand must be alloca or constant.", &CPI);
+
   visitEHPadPredecessors(CPI);
   visitFuncletPadInst(CPI);
 }
diff --git a/llvm/lib/Transforms/Coroutines/SpillUtils.cpp b/llvm/lib/Transforms/Coroutines/SpillUtils.cpp
index 81fe0c9acd413..05abccf0f9a97 100644
--- a/llvm/lib/Transforms/Coroutines/SpillUtils.cpp
+++ b/llvm/lib/Transforms/Coroutines/SpillUtils.cpp
@@ -180,6 +180,13 @@ struct AllocaUseVisitor : PtrUseVisitor<AllocaUseVisitor> {
     handleAlias(I);
   }
 
+  void visitCatchPadInst(CatchPadInst &I) {
+    // Windows EH requires exception objects allocated on the stack,
+    // shortcut the traversal and keep it on stack.
+    ShouldLiveOnFrame = false;
+    Base::Worklist.clear();
+  }
+
   void visitInsertElementInst(InsertElementInst &I) {
     enqueueUsers(I);
     handleAlias(I);
diff --git a/llvm/test/Transforms/Coroutines/coro-alloca-10.ll b/llvm/test/Transforms/Coroutines/coro-alloca-10.ll
new file mode 100644
index 0000000000000..a5fb67d59edd1
--- /dev/null
+++ b/llvm/test/Transforms/Coroutines/coro-alloca-10.ll
@@ -0,0 +1,66 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; Test that catchpad is specially handled. Do not collect exception object into coroutine frame.
+; RUN: opt < %s -passes='coro-split,simplifycfg,early-cse' -S | FileCheck %s
+
+define void @fn() presplitcoroutine personality i32 0 {
+; CHECK-LABEL: define void @fn() personality i32 0 {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[EXCEPTION_OBJ_RELOAD_ADDR:%.*]] = alloca ptr, align 8
+; CHECK-NEXT:    [[ID:%.*]] = call token @llvm.coro.id(i32 16, ptr null, ptr null, ptr @fn.resumers)
+; CHECK-NEXT:    [[MEM:%.*]] = call noalias nonnull ptr @malloc(i64 24)
+; CHECK-NEXT:    [[HDL:%.*]] = call noalias nonnull ptr @llvm.coro.begin(token [[ID]], ptr [[MEM]])
+; CHECK-NEXT:    store ptr @fn.resume, ptr [[HDL]], align 8
+; CHECK-NEXT:    [[DESTROY_ADDR:%.*]] = getelementptr inbounds nuw [[FN_FRAME:%.*]], ptr [[HDL]], i32 0, i32 1
+; CHECK-NEXT:    store ptr @fn.destroy, ptr [[DESTROY_ADDR]], align 8
+; CHECK-NEXT:    store ptr null, ptr [[EXCEPTION_OBJ_RELOAD_ADDR]], align 8
+; CHECK-NEXT:    [[INDEX_ADDR4:%.*]] = getelementptr inbounds nuw [[FN_FRAME]], ptr [[HDL]], i32 0, i32 2
+; CHECK-NEXT:    store i1 false, ptr [[INDEX_ADDR4]], align 1
+; CHECK-NEXT:    ret void
+;
+entry:
+  %exception.obj = alloca ptr, align 8
+  %id = call token @llvm.coro.id(i32 16, ptr null, ptr null, ptr null)
+  %size = call i64 @llvm.coro.size.i64()
+  %mem = call noalias nonnull ptr @malloc(i64 %size)
+  %hdl = call ptr @llvm.coro.begin(token %id, ptr %mem)
+  store ptr null, ptr %exception.obj, align 8
+  br label %while
+
+while:
+  %save = call token @llvm.coro.save(ptr null)
+  %suspend = call i8 @llvm.coro.suspend(token %save, i1 false)
+  switch i8 %suspend, label %coro.ret [
+  i8 0, label %await.ready
+  ]
+
+await.ready:
+  invoke void @throw()
+  to label %unreachable unwind label %catch.dispatch
+
+catch.dispatch:
+  %switch = catchswitch within none [label %catch] unwind label %ehcleanup
+
+catch:
+  %pad = catchpad within %switch [ptr null, i32 8, ptr %exception.obj]
+  invoke void @use(ptr %exception.obj) [ "funclet"(token %pad) ]
+  to label %catch.ret unwind label %ehcleanup
+
+catch.ret:
+  catchret from %pad to label %while
+
+ehcleanup:
+  %cleanup = cleanuppad within none []
+  call void @llvm.coro.end(ptr null, i1 true, token none) [ "funclet"(token %cleanup) ]
+  cleanupret from %cleanup unwind to caller
+
+coro.ret:
+  call void @llvm.coro.end(ptr null, i1 false, token none)
+  ret void
+
+unreachable:
+  unreachable
+}
+
+declare ptr @malloc(i64)
+declare void @throw()
+declare void @use(ptr)
diff --git a/llvm/unittests/FuzzMutate/StrategiesTest.cpp b/llvm/unittests/FuzzMutate/StrategiesTest.cpp
index fd91066b994a5..278cce20119c6 100644
--- a/llvm/unittests/FuzzMutate/StrategiesTest.cpp
+++ b/llvm/unittests/FuzzMutate/StrategiesTest.cpp
@@ -733,11 +733,12 @@ TEST(AllStrategies, SkipEHPad) {
   StringRef Source = "\n\
     define void @f(i32 %x) personality ptr @__CxxFrameHandler3 { \n\
     entry: \n\
+      %I = alloca i32, align 4 \n\
       invoke void @g() to label %try.cont unwind label %catch.dispatch \n\
     catch.dispatch: \n\
       %0 = catchswitch within none [label %catch] unwind to caller \n\
     catch: \n\
-      %1 = catchpad within %0 [ptr null, i32 64, ptr null] \n\
+      %1 = catchpad within %0 [ptr null, i32 64, ptr %I] \n\
       catchret from %1 to label %try.cont \n\
     try.cont: \n\
       ret void \n\

Comment thread llvm/lib/IR/Verifier.cpp Outdated
auto *V = U.get();
return isa<Constant>(V) || isa<AllocaInst>(V);
}),
"Argument operand must be alloca or constant.", &CPI);

@nikic nikic Apr 30, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm concerned about backporting changes to the IR verifier.

@NewSigma NewSigma closed this Apr 30, 2026
@NewSigma NewSigma removed this from the LLVM 22.x Release milestone Apr 30, 2026
@nikic

nikic commented Apr 30, 2026

Copy link
Copy Markdown
Contributor

Just to be clear, I think backporting just the SpillUtils change would probably be fine. I think that one doesn't strictly depend on the LangRef/Verifier changes.

@NewSigma NewSigma reopened this Apr 30, 2026
@github-project-automation github-project-automation Bot moved this from Needs Review to Needs Triage in LLVM Release Status Apr 30, 2026
@NewSigma

Copy link
Copy Markdown
Contributor Author

Just to be clear, I think backporting just the SpillUtils change would probably be fine. I think that one doesn't strictly depend on the LangRef/Verifier changes.

Oh, I closed it too quickly. The verifier’s changes have been reverted.

@NewSigma NewSigma added this to the LLVM 22.x Release milestone Apr 30, 2026
@NewSigma NewSigma removed the llvm:ir label Apr 30, 2026
@dyung dyung moved this from Needs Triage to Needs Review in LLVM Release Status May 1, 2026
@dyung

dyung commented May 1, 2026

Copy link
Copy Markdown
Contributor

Is this a regression from a previous release? If so, which release?

@NewSigma

NewSigma commented May 1, 2026

Copy link
Copy Markdown
Contributor Author

Is this a regression from a previous release? If so, which release?

No, this fixes a bug that has been present since the first release that supported coroutines.

@dyung

dyung commented May 4, 2026

Copy link
Copy Markdown
Contributor

Is this a regression from a previous release? If so, which release?

No, this fixes a bug that has been present since the first release that supported coroutines.

What are the implications if we do not fix this bug in the 22.x release? The change seems small which is good, but if this is a serious enough issue to warrant inclusion in the 22.x branch.

@NewSigma

NewSigma commented May 4, 2026

Copy link
Copy Markdown
Contributor Author

Is this a regression from a previous release? If so, which release?

No, this fixes a bug that has been present since the first release that supported coroutines.

What are the implications if we do not fix this bug in the 22.x release? The change seems small which is good, but if this is a serious enough issue to warrant inclusion in the 22.x branch.

If this bug is not fixed in the 22.x release, Windows users will continue to experience crashes. Given that there are several related issues and a request to backport it, this may be an important fix worth including. I have no strong opinion and will respect release management’s decision.

@agarcin

agarcin commented May 4, 2026

Copy link
Copy Markdown

Is this a regression from a previous release? If so, which release?

No, this fixes a bug that has been present since the first release that supported coroutines.

What are the implications if we do not fix this bug in the 22.x release? The change seems small which is good, but if this is a serious enough issue to warrant inclusion in the 22.x branch.

For my part, without this fix I’ve had to comment out every section of my code that uses the caught exception object. It works as a temporary workaround since I only rely on e.what() for debugging, but it’s still inconvenient — those parts must stay commented out, so I can’t actually use the exception object at the moment.

I’d be really grateful to see this addressed in a clang‑cl release, as it would remove that friction entirely for Windows users like me.
And thanks again for the PR — really appreciate the work you’re putting into this.

@dyung

dyung commented May 14, 2026

Copy link
Copy Markdown
Contributor

If this change is to go in, it needs a reviewer to approve it.

@github-project-automation github-project-automation Bot moved this from Needs Review to Needs Merge in LLVM Release Status May 14, 2026
@dyung

dyung commented May 14, 2026

Copy link
Copy Markdown
Contributor

@NewSigma I've squashed your 3 commits into 1, can you do a quick review to make sure it all still looks correct?

@NewSigma

Copy link
Copy Markdown
Contributor Author

@NewSigma I've squashed your 3 commits into 1, can you do a quick review to make sure it all still looks correct?

LG

…86728)

Windows EH requires exception objects allocated on stack. But there is
no reliable way to identify them. CoroSplit employs a best-effort
algorithm to determine whether allocas persist on the stack or the
frame, which may result in miscompilation when Windows exceptions are
used.
This patch proposes that we treat allocas used by catchpad as exception
objects and never place them on the frame. A verifier check is added to
enforce that operands of catchpad are either constants or allocas.

Close llvm#143235 Close llvm#153949 Close llvm#182584
@c-rhodes
c-rhodes merged commit e624f12 into llvm:release/22.x May 26, 2026
@github-project-automation github-project-automation Bot moved this from Needs Merge to Done in LLVM Release Status May 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

Development

Successfully merging this pull request may close these issues.

7 participants