1 files changed, 2 insertions, 1 deletions

diff --git a/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp b/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
index c32303defe7f..015c71080d67 100644
--- a/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
+++ b/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
@@ -106,7 +106,7 @@ static bool lowerKernelArguments(Function &F, const TargetMachine &TM) {
   LLVMContext &Ctx = F.getParent()->getContext();
   const DataLayout &DL = F.getParent()->getDataLayout();
   BasicBlock &EntryBlock = *F.begin();
-  IRBuilder<> Builder(&*getInsertPt(EntryBlock));
+  IRBuilder<> Builder(&EntryBlock, getInsertPt(EntryBlock));
 
   const Align KernArgBaseAlign(16); // FIXME: Increase if necessary
   const uint64_t BaseOffset = ST.getExplicitKernelArgOffset();
@@ -202,6 +202,7 @@ static bool lowerKernelArguments(Function &F, const TargetMachine &TM) {
       // Since we don't have sub-dword scalar loads, avoid doing an extload by
       // loading earlier than the argument address, and extracting the relevant
       // bits.
+      // TODO: Update this for GFX12 which does have scalar sub-dword loads.
       //
       // Additionally widen any sub-dword load to i32 even if suitably aligned,
       // so that CSE between different argument loads works easily.