[Clang][CodeGen] Don't emit assumptions if current block is unreachable.

[dtcxzyw]

Fixes #106898.

When emitting an infinite loop, clang codegen will delete the whole block and leave builder's current block as nullptr:

llvm-project/clang/lib/CodeGen/CGStmt.cpp

Lines 597 to 600 in 837ee5b

	if (IsFinished && BB->use_empty()) {
	delete BB;
	return;
	}

Then clang will create zext (icmp slt %a, %b) without parent block for a < b. It will crash here:

llvm-project/clang/lib/CodeGen/CGExprScalar.cpp

Lines 416 to 420 in 837ee5b

	// If there aren't any more uses, zap the instruction to save space.
	// Note that there can be more uses, for example if this
	// is the result of an assignment.
	if (ZI->use_empty())
	ZI->eraseFromParent();

Even if we disabled this optimization, it still crashes in Builder.CreateAssumption:

llvm-project/llvm/lib/IR/IRBuilder.cpp

Lines 551 to 561 in 837ee5b

	CallInst *
	IRBuilderBase::CreateAssumption(Value *Cond,
	ArrayRef<OperandBundleDef> OpBundles) {
	assert(Cond->getType() == getInt1Ty() &&
	"an assumption condition must be of type i1");
	Value *Ops[] = { Cond };
	Module *M = BB->getParent()->getParent();
	Function *FnAssume = Intrinsic::getDeclaration(M, Intrinsic::assume);
	return CreateCall(FnAssume, Ops, OpBundles);
	}

This patch disables assumptions emission if current block is null. As an alternative, we can fix the optimization in EmitIntToBoolConversion and use CGM.getIntrinsic as we do for __builtin_assume:

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

Lines 3422 to 3431 in 837ee5b

	case Builtin::BI__assume:
	case Builtin::BI__builtin_assume: {
	if (E->getArg(0)->HasSideEffects(getContext()))
	return RValue::get(nullptr);
	Value *ArgValue = EmitScalarExpr(E->getArg(0));
	Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
	Builder.CreateCall(FnAssume, ArgValue);
	return RValue::get(nullptr);
	}

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-codegen

Author: Yingwei Zheng (dtcxzyw)

Changes

Fixes #106898.

When emitting an infinite loop, clang codegen will delete the whole block and leave builder's current block as nullptr:

llvm-project/clang/lib/CodeGen/CGStmt.cpp

Lines 597 to 600 in 837ee5b

	if (IsFinished && BB->use_empty()) {
	delete BB;
	return;
	}

Then clang will create zext (icmp slt %a, %b) without parent block for a < b. It will crash here:

llvm-project/clang/lib/CodeGen/CGExprScalar.cpp

Lines 416 to 420 in 837ee5b

	// If there aren't any more uses, zap the instruction to save space.
	// Note that there can be more uses, for example if this
	// is the result of an assignment.
	if (ZI->use_empty())
	ZI->eraseFromParent();

Even if we disabled this optimization, it still crashes in Builder.CreateAssumption:

llvm-project/llvm/lib/IR/IRBuilder.cpp

Lines 551 to 561 in 837ee5b

	CallInst *
	IRBuilderBase::CreateAssumption(Value *Cond,
	ArrayRef<OperandBundleDef> OpBundles) {
	assert(Cond->getType() == getInt1Ty() &&
	"an assumption condition must be of type i1");
	Value *Ops[] = { Cond };
	Module *M = BB->getParent()->getParent();
	Function *FnAssume = Intrinsic::getDeclaration(M, Intrinsic::assume);
	return CreateCall(FnAssume, Ops, OpBundles);
	}

This patch disables assumptions emission if current block is null. As an alternative, we can fix the optimization in EmitIntToBoolConversion and use CGM.getIntrinsic as we do for __builtin_assume:

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

Lines 3422 to 3431 in 837ee5b

	case Builtin::BI__assume:
	case Builtin::BI__builtin_assume: {
	if (E->getArg(0)->HasSideEffects(getContext()))
	return RValue::get(nullptr);
	Value *ArgValue = EmitScalarExpr(E->getArg(0));
	Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
	Builder.CreateCall(FnAssume, ArgValue);
	return RValue::get(nullptr);
	}

---

Full diff: https://github.com/llvm/llvm-project/pull/106936.diff

2 Files Affected:

• (modified) clang/lib/CodeGen/CGStmt.cpp (+1-1)
• (modified) clang/test/SemaCXX/cxx23-assume.cpp (+9)

diff --git a/clang/lib/CodeGen/CGStmt.cpp b/clang/lib/CodeGen/CGStmt.cpp
index 7158a06e6bc3b3..b138c87a853495 100644
--- a/clang/lib/CodeGen/CGStmt.cpp
+++ b/clang/lib/CodeGen/CGStmt.cpp
@@ -752,7 +752,7 @@ void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
     } break;
     case attr::CXXAssume: {
       const Expr *Assumption = cast<CXXAssumeAttr>(A)->getAssumption();
-      if (getLangOpts().CXXAssumptions &&
+      if (getLangOpts().CXXAssumptions && Builder.GetInsertBlock() &&
           !Assumption->HasSideEffects(getContext())) {
         llvm::Value *AssumptionVal = EvaluateExprAsBool(Assumption);
         Builder.CreateAssumption(AssumptionVal);
diff --git a/clang/test/SemaCXX/cxx23-assume.cpp b/clang/test/SemaCXX/cxx23-assume.cpp
index 9138501d726dd6..eeae59daea3f70 100644
--- a/clang/test/SemaCXX/cxx23-assume.cpp
+++ b/clang/test/SemaCXX/cxx23-assume.cpp
@@ -158,3 +158,12 @@ foo (int x, int y)
   return x + y;
 }
 }
+
+// Do not crash when assumptions are unreachable.
+namespace gh106898 {
+int foo () { 
+    while(1);
+    int a = 0, b = 1;
+    __attribute__((assume (a < b)));
+}
+}

[efriedma-quic]

LGTM

[dtcxzyw]

/cherry-pick c94bd96

[llvmbot]

/pull-request #107183