aboutsummaryrefslogtreecommitdiff
path: root/lib/CodeGen/CGException.cpp
diff options
context:
space:
mode:
authorJohn McCall <rjmccall@apple.com>2011-05-27 20:01:14 +0000
committerJohn McCall <rjmccall@apple.com>2011-05-27 20:01:14 +0000
commit0903421e36c174a82597f83bd296f3cd5b5f169b (patch)
treea39425c04659ed5fa2de07f8633473f4e632ee66 /lib/CodeGen/CGException.cpp
parentf996e051d9953550982b57132daad8a5e3f7bd65 (diff)
Implement a new, much improved version of the cleanup hack. We just need
to be careful to emit landing pads that are always prepared to handle a cleanup path. This is correct mostly because of the fix to the LLVM inliner, r132200. git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@132209 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/CodeGen/CGException.cpp')
-rw-r--r--lib/CodeGen/CGException.cpp131
1 files changed, 80 insertions, 51 deletions
diff --git a/lib/CodeGen/CGException.cpp b/lib/CodeGen/CGException.cpp
index 6cb9599e25..91be321d31 100644
--- a/lib/CodeGen/CGException.cpp
+++ b/lib/CodeGen/CGException.cpp
@@ -112,11 +112,18 @@ static llvm::Constant *getUnexpectedFn(CodeGenFunction &CGF) {
return CGF.CGM.CreateRuntimeFunction(FTy, "__cxa_call_unexpected");
}
+llvm::Constant *CodeGenFunction::getUnwindResumeFn() {
+ const llvm::FunctionType *FTy =
+ llvm::FunctionType::get(VoidTy, Int8PtrTy, /*IsVarArgs=*/false);
+
+ if (CGM.getLangOptions().SjLjExceptions)
+ return CGM.CreateRuntimeFunction(FTy, "_Unwind_SjLj_Resume");
+ return CGM.CreateRuntimeFunction(FTy, "_Unwind_Resume");
+}
+
llvm::Constant *CodeGenFunction::getUnwindResumeOrRethrowFn() {
- const llvm::Type *Int8PtrTy = llvm::Type::getInt8PtrTy(getLLVMContext());
const llvm::FunctionType *FTy =
- llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), Int8PtrTy,
- /*IsVarArgs=*/false);
+ llvm::FunctionType::get(VoidTy, Int8PtrTy, /*IsVarArgs=*/false);
if (CGM.getLangOptions().SjLjExceptions)
return CGM.CreateRuntimeFunction(FTy, "_Unwind_SjLj_Resume_or_Rethrow");
@@ -563,47 +570,59 @@ llvm::BasicBlock *CodeGenFunction::getInvokeDestImpl() {
return LP;
}
+// This code contains a hack to work around a design flaw in
+// LLVM's EH IR which breaks semantics after inlining. This same
+// hack is implemented in llvm-gcc.
+//
+// The LLVM EH abstraction is basically a thin veneer over the
+// traditional GCC zero-cost design: for each range of instructions
+// in the function, there is (at most) one "landing pad" with an
+// associated chain of EH actions. A language-specific personality
+// function interprets this chain of actions and (1) decides whether
+// or not to resume execution at the landing pad and (2) if so,
+// provides an integer indicating why it's stopping. In LLVM IR,
+// the association of a landing pad with a range of instructions is
+// achieved via an invoke instruction, the chain of actions becomes
+// the arguments to the @llvm.eh.selector call, and the selector
+// call returns the integer indicator. Other than the required
+// presence of two intrinsic function calls in the landing pad,
+// the IR exactly describes the layout of the output code.
+//
+// A principal advantage of this design is that it is completely
+// language-agnostic; in theory, the LLVM optimizers can treat
+// landing pads neutrally, and targets need only know how to lower
+// the intrinsics to have a functioning exceptions system (assuming
+// that platform exceptions follow something approximately like the
+// GCC design). Unfortunately, landing pads cannot be combined in a
+// language-agnostic way: given selectors A and B, there is no way
+// to make a single landing pad which faithfully represents the
+// semantics of propagating an exception first through A, then
+// through B, without knowing how the personality will interpret the
+// (lowered form of the) selectors. This means that inlining has no
+// choice but to crudely chain invokes (i.e., to ignore invokes in
+// the inlined function, but to turn all unwindable calls into
+// invokes), which is only semantically valid if every unwind stops
+// at every landing pad.
+//
+// Therefore, the invoke-inline hack is to guarantee that every
+// landing pad has a catch-all.
+enum CleanupHackLevel_t {
+ /// A level of hack that requires that all landing pads have
+ /// catch-alls.
+ CHL_MandatoryCatchall,
+
+ /// A level of hack that requires that all landing pads handle
+ /// cleanups.
+ CHL_MandatoryCleanup,
+
+ /// No hacks at all; ideal IR generation.
+ CHL_Ideal
+};
+const CleanupHackLevel_t CleanupHackLevel = CHL_MandatoryCleanup;
+
llvm::BasicBlock *CodeGenFunction::EmitLandingPad() {
assert(EHStack.requiresLandingPad());
- // This function contains a hack to work around a design flaw in
- // LLVM's EH IR which breaks semantics after inlining. This same
- // hack is implemented in llvm-gcc.
- //
- // The LLVM EH abstraction is basically a thin veneer over the
- // traditional GCC zero-cost design: for each range of instructions
- // in the function, there is (at most) one "landing pad" with an
- // associated chain of EH actions. A language-specific personality
- // function interprets this chain of actions and (1) decides whether
- // or not to resume execution at the landing pad and (2) if so,
- // provides an integer indicating why it's stopping. In LLVM IR,
- // the association of a landing pad with a range of instructions is
- // achieved via an invoke instruction, the chain of actions becomes
- // the arguments to the @llvm.eh.selector call, and the selector
- // call returns the integer indicator. Other than the required
- // presence of two intrinsic function calls in the landing pad,
- // the IR exactly describes the layout of the output code.
- //
- // A principal advantage of this design is that it is completely
- // language-agnostic; in theory, the LLVM optimizers can treat
- // landing pads neutrally, and targets need only know how to lower
- // the intrinsics to have a functioning exceptions system (assuming
- // that platform exceptions follow something approximately like the
- // GCC design). Unfortunately, landing pads cannot be combined in a
- // language-agnostic way: given selectors A and B, there is no way
- // to make a single landing pad which faithfully represents the
- // semantics of propagating an exception first through A, then
- // through B, without knowing how the personality will interpret the
- // (lowered form of the) selectors. This means that inlining has no
- // choice but to crudely chain invokes (i.e., to ignore invokes in
- // the inlined function, but to turn all unwindable calls into
- // invokes), which is only semantically valid if every unwind stops
- // at every landing pad.
- //
- // Therefore, the invoke-inline hack is to guarantee that every
- // landing pad has a catch-all.
- const bool UseInvokeInlineHack = true;
-
for (EHScopeStack::iterator ir = EHStack.begin(); ; ) {
assert(ir != EHStack.end() &&
"stack requiring landing pad is nothing but non-EH scopes?");
@@ -736,16 +755,23 @@ llvm::BasicBlock *CodeGenFunction::EmitLandingPad() {
EHSelector.append(EHFilters.begin(), EHFilters.end());
// Also check whether we need a cleanup.
- if (UseInvokeInlineHack || HasEHCleanup)
- EHSelector.push_back(UseInvokeInlineHack
+ if (CleanupHackLevel == CHL_MandatoryCatchall || HasEHCleanup)
+ EHSelector.push_back(CleanupHackLevel == CHL_MandatoryCatchall
? getCatchAllValue(*this)
: getCleanupValue(*this));
// Otherwise, signal that we at least have cleanups.
- } else if (UseInvokeInlineHack || HasEHCleanup) {
- EHSelector.push_back(UseInvokeInlineHack
+ } else if (CleanupHackLevel == CHL_MandatoryCatchall || HasEHCleanup) {
+ EHSelector.push_back(CleanupHackLevel == CHL_MandatoryCatchall
? getCatchAllValue(*this)
: getCleanupValue(*this));
+
+ // At the MandatoryCleanup hack level, we don't need to actually
+ // spuriously tell the unwinder that we have cleanups, but we do
+ // need to always be prepared to handle cleanups.
+ } else if (CleanupHackLevel == CHL_MandatoryCleanup) {
+ // Just don't decrement LastToEmitInLoop.
+
} else {
assert(LastToEmitInLoop > 2);
LastToEmitInLoop--;
@@ -833,7 +859,7 @@ llvm::BasicBlock *CodeGenFunction::EmitLandingPad() {
// If there was a cleanup, we'll need to actually check whether we
// landed here because the filter triggered.
- if (UseInvokeInlineHack || HasEHCleanup) {
+ if (CleanupHackLevel != CHL_Ideal || HasEHCleanup) {
llvm::BasicBlock *RethrowBB = createBasicBlock("cleanup");
llvm::BasicBlock *UnexpectedBB = createBasicBlock("ehspec.unexpected");
@@ -843,10 +869,11 @@ llvm::BasicBlock *CodeGenFunction::EmitLandingPad() {
Builder.CreateCondBr(FailsFilter, UnexpectedBB, RethrowBB);
// The rethrow block is where we land if this was a cleanup.
- // TODO: can this be _Unwind_Resume if the InvokeInlineHack is off?
EmitBlock(RethrowBB);
- Builder.CreateCall(getUnwindResumeOrRethrowFn(),
- Builder.CreateLoad(getExceptionSlot()))
+ llvm::Constant *RethrowFn =
+ CleanupHackLevel == CHL_MandatoryCatchall ? getUnwindResumeOrRethrowFn()
+ : getUnwindResumeFn();
+ Builder.CreateCall(RethrowFn, Builder.CreateLoad(getExceptionSlot()))
->setDoesNotReturn();
Builder.CreateUnreachable();
@@ -863,7 +890,7 @@ llvm::BasicBlock *CodeGenFunction::EmitLandingPad() {
Builder.CreateUnreachable();
// ...or a normal catch handler...
- } else if (!UseInvokeInlineHack && !HasEHCleanup) {
+ } else if (CleanupHackLevel == CHL_Ideal && !HasEHCleanup) {
llvm::Value *Type = EHSelector.back();
EmitBranchThroughEHCleanup(EHHandlers[Type]);
@@ -1444,7 +1471,9 @@ CodeGenFunction::UnwindDest CodeGenFunction::getRethrowDest() {
if (!RethrowName.empty())
RethrowFn = getCatchallRethrowFn(*this, RethrowName);
else
- RethrowFn = getUnwindResumeOrRethrowFn();
+ RethrowFn = (CleanupHackLevel == CHL_MandatoryCatchall
+ ? getUnwindResumeOrRethrowFn()
+ : getUnwindResumeFn());
Builder.CreateCall(RethrowFn, Builder.CreateLoad(getExceptionSlot()))
->setDoesNotReturn();