[AMDGPU] Limit promoting allocas that have users with dynamic index above a threshold on number of elements

llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

@@ -85,6 +85,12 @@ static cl::opt<unsigned>
                             "when sorting profitable allocas"),
                    cl::init(4));
 
+static cl::opt<unsigned> PromoteAllocaDynamicIndexNumberElementLimit(
+    "amdgpu-promote-alloca-dynamic-index-num-element-limit",
+    cl::desc("Maximum number of elements for promoting alloca with dynamic"
+             " index"),
+    cl::init(8));
+
 // Shared implementation which can do both promotion to vector and to LDS.
 class AMDGPUPromoteAllocaImpl {
 private:
@@ -920,6 +926,19 @@ bool AMDGPUPromoteAllocaImpl::tryPromoteAllocaToVector(AllocaInst &Alloca) {
       if (!Index)
         return RejectUser(Inst, "cannot compute vector index for GEP");
 
+      if (!isa<ConstantInt>(Index)) {
+        bool UsedInLoad = llvm::any_of(
+            GEP->users(), [&](const auto *U) { return isa<LoadInst>(U); });
+        if (auto *UserVecTy =
+                dyn_cast<FixedVectorType>(GEP->getSourceElementType())) {
+          if (UsedInLoad && UserVecTy->getNumElements() >
+                                PromoteAllocaDynamicIndexNumberElementLimit) {
+            return RejectUser(Inst,
+                              "user has too many elements for dynamic index");
+          }
+        }
+      }
+
       GEPVectorIdx[GEP] = Index;
       UsersToRemove.push_back(Inst);
       continue;