[GISel] Check useLoadStackGuardNode() before generating LOAD_STACK_GUARD

[Kai]

Summary

When lowering llvm::stackprotect intrinsic, the SDAG implementation
checks function useLoadStackGuardNode() to either create a
LOAD_STACK_GUARD node or use the first argument of the intrinsic.
This check is not present in the IRTranslator, which results in always
generating a LOAD_STACK_GUARD even if the target does not support it.

[arsenm]

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2083

Don't see why getStackGuard doesn't just return the virtual register it creates itself

llvm/test/CodeGen/PowerPC/GlobalISel/ppc-irtranslator-stackprotect.ll

3

Is this actually setting the triple correctly? The OS should be in the 3rd position?

[Kai]

Fixed triple.

[Kai]

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2083

Sorry, do you mean that getStackGuard() should create the virtual register? Currently it uses the register passed as argument.
I guess that should work as LLVM always uses a pointer as type for the stack guard value.

[arsenm]

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2083

Yes. These APIs that take output registers are a bit awkward

[Kai]

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

2083

That is not as trivial as it seems.
In emitSPDescriptorParent(), the type of the guard value is deliberately changed to a scalar similar in size of a pointer, to match the type of the G_LOAD instruction loading the guard value from the stack slot.
That is very nicely demonstrated in test case AArch64/GlobalISel/irtranslator-delayed-stack-protector.ll.
I can make this change but I would prefer to do it in a separate change, as it is really a different goal ("unifying use of getStackGuard()").

llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp

bool IRTranslator::translateKnownIntrinsic(const CallInst &CI, Intrinsic::ID ID,
 
   case Intrinsic::is_constant:
     llvm_unreachable("llvm.is.constant.* should have been lowered already");
 
   case Intrinsic::stackguard:
     getStackGuard(getOrCreateVReg(CI), MIRBuilder);
     return true;
   case Intrinsic::stackprotector: {
+    const TargetLowering &TLI = *MF->getSubtarget().getTargetLowering();
     LLT PtrTy = getLLTForType(*CI.getArgOperand(0)->getType(), *DL);
-    Register GuardVal = MRI->createGenericVirtualRegister(PtrTy);
+    Register GuardVal;
+    if (TLI.useLoadStackGuardNode()) {
+      GuardVal = MRI->createGenericVirtualRegister(PtrTy);
       getStackGuard(GuardVal, MIRBuilder);
+    } else
+      GuardVal = getOrCreateVReg(*CI.getArgOperand(0)); // The guard's value.
 
     AllocaInst *Slot = cast<AllocaInst>(CI.getArgOperand(1));
     int FI = getOrCreateFrameIndex(*Slot);
     MF->getFrameInfo().setStackProtectorIndex(FI);
 
     MIRBuilder.buildStore(
         GuardVal, getOrCreateVReg(*Slot),
         *MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(*MF, FI),

llvm/test/CodeGen/PowerPC/GlobalISel/ppc-irtranslator-stackprotect.ll

+; RUN: llc -verify-machineinstrs -mtriple=ppc64le-unknown-linux %s -global-isel -stop-after=irtranslator -o - | FileCheck %s --check-prefix=LINUX
+; RUN: llc -verify-machineinstrs -mtriple=ppc64le-unknown-openbsd %s -global-isel -stop-after=irtranslator -o - | FileCheck %s --check-prefix=OPENBSD
+
+
+; The stack guard on Linux
+@__stack_chk_guard = external global ptr
+
+; The stack guard on OpenBSD
+@__guard_local = external hidden global ptr
+
+declare void @llvm.stackprotector(ptr, ptr)
+
+; LINUX-LABEL: name: test_stack_guard_linux
+
+; LINUX: frameInfo:
+; LINUX: stackProtector:  '%stack.0.StackGuardSlot'
+
+; LINUX: stack:
+; LINUX:  - { id: 0, name: StackGuardSlot,  type: default, offset: 0, size: 8, alignment: 8,
+; LINUX-NOT: id: 1
+
+; LINUX: [[GUARD_SLOT:%[0-9]+]]:_(p0) = G_FRAME_INDEX %stack.0.StackGuardSlot
+; LINUX: [[GUARD:%[0-9]+]]:g8rc(p0) = LOAD_STACK_GUARD :: (dereferenceable invariant load (p0) from @__stack_chk_guard)
+; LINUX: G_STORE [[GUARD]](p0), [[GUARD_SLOT]](p0) :: (volatile store (p0) into %stack.0.StackGuardSlot)
+define void @test_stack_guard_linux() {
+  %StackGuardSlot = alloca ptr
+  call void @llvm.stackprotector(ptr undef, ptr %StackGuardSlot)
+  ret void
+}
+
+; OPENBSD-LABEL: name: test_stack_guard_openbsd
+
+; OPENBSD: frameInfo:
+; OPENBSD: stackProtector:  '%stack.0.StackGuardSlot'
+
+; OPENBSD: stack:
+; OPENBSD:  - { id: 0, name: StackGuardSlot,  type: default, offset: 0, size: 8, alignment: 8,
+; OPENBSD-NOT: id: 1
+
+; OPENBSD: [[GUARD_LOCAL:%[0-9]+]]:_(p0) = G_GLOBAL_VALUE @__guard_local
+; OPENBSD: [[GUARD_SLOT:%[0-9]+]]:_(p0) = G_FRAME_INDEX %stack.0.StackGuardSlot
+; OPENBSD: [[GUARD:%[0-9]+]]:_(p0) = G_LOAD [[GUARD_LOCAL]](p0) :: (dereferenceable load (p0) from @__guard_local)
+; OPENBSD: G_STORE [[GUARD]](p0), [[GUARD_SLOT]](p0) :: (volatile store (p0) into %stack.0.StackGuardSlot)
+define void @test_stack_guard_openbsd() {
+  %StackGuardSlot = alloca i8*
+  %StackGuard = load ptr, ptr @__guard_local
+  call void @llvm.stackprotector(ptr %StackGuard, ptr %StackGuardSlot)
+  ret void
+}