1 files changed, 21 insertions, 6 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
index e64542a395f0..62ab5bb55a16 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp
@@ -58,6 +58,21 @@ public:
 
 } // end anonymous namespace
 
+// skip allocas
+static BasicBlock::iterator getInsertPt(BasicBlock &BB) {
+  BasicBlock::iterator InsPt = BB.getFirstInsertionPt();
+  for (BasicBlock::iterator E = BB.end(); InsPt != E; ++InsPt) {
+    AllocaInst *AI = dyn_cast<AllocaInst>(&*InsPt);
+
+    // If this is a dynamic alloca, the value may depend on the loaded kernargs,
+    // so loads will need to be inserted before it.
+    if (!AI || !AI->isStaticAlloca())
+      break;
+  }
+
+  return InsPt;
+}
+
 bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
   CallingConv::ID CC = F.getCallingConv();
   if (CC != CallingConv::AMDGPU_KERNEL || F.arg_empty())
@@ -70,7 +85,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
   LLVMContext &Ctx = F.getParent()->getContext();
   const DataLayout &DL = F.getParent()->getDataLayout();
   BasicBlock &EntryBlock = *F.begin();
-  IRBuilder<> Builder(&*EntryBlock.begin());
+  IRBuilder<> Builder(&*getInsertPt(EntryBlock));
 
   const Align KernArgBaseAlign(16); // FIXME: Increase if necessary
   const uint64_t BaseOffset = ST.getExplicitKernelArgOffset(F);
@@ -94,7 +109,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
 
   for (Argument &Arg : F.args()) {
     Type *ArgTy = Arg.getType();
-    unsigned ABITypeAlign = DL.getABITypeAlignment(ArgTy);
+    Align ABITypeAlign = DL.getABITypeAlign(ArgTy);
     unsigned Size = DL.getTypeSizeInBits(ArgTy);
     unsigned AllocSize = DL.getTypeAllocSize(ArgTy);
 
@@ -120,7 +135,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
         continue;
     }
 
-    VectorType *VT = dyn_cast<VectorType>(ArgTy);
+    auto *VT = dyn_cast<FixedVectorType>(ArgTy);
     bool IsV3 = VT && VT->getNumElements() == 3;
     bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType();
 
@@ -152,7 +167,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
     }
 
     if (IsV3 && Size >= 32) {
-      V4Ty = VectorType::get(VT->getVectorElementType(), 4);
+      V4Ty = FixedVectorType::get(VT->getElementType(), 4);
       // Use the hack that clang uses to avoid SelectionDAG ruining v3 loads
       AdjustedArgTy = V4Ty;
     }
@@ -160,7 +175,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
     ArgPtr = Builder.CreateBitCast(ArgPtr, AdjustedArgTy->getPointerTo(AS),
                                    ArgPtr->getName() + ".cast");
     LoadInst *Load =
-        Builder.CreateAlignedLoad(AdjustedArgTy, ArgPtr, AdjustedAlign.value());
+        Builder.CreateAlignedLoad(AdjustedArgTy, ArgPtr, AdjustedAlign);
     Load->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(Ctx, {}));
 
     MDBuilder MDB(Ctx);
@@ -210,7 +225,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
       Arg.replaceAllUsesWith(NewVal);
     } else if (IsV3) {
       Value *Shuf = Builder.CreateShuffleVector(Load, UndefValue::get(V4Ty),
-                                                {0, 1, 2},
+                                                ArrayRef<int>{0, 1, 2},
                                                 Arg.getName() + ".load");
       Arg.replaceAllUsesWith(Shuf);
     } else {