diff --git a/src/codegen/llvm/codegen_amdgpu.cc b/src/codegen/llvm/codegen_amdgpu.cc
index fdd2ec2e38d24a1287552f90e534352cf93e9255..9b8995bf5516ef1d1ab93a4c52d12b9c6c2a7176 100644
--- a/src/codegen/llvm/codegen_amdgpu.cc
+++ b/src/codegen/llvm/codegen_amdgpu.cc
@@ -176,6 +176,9 @@ runtime::Module BuildAMDGPU(Array<LoweredFunc> funcs, std::string target) {
     }
     mlib->setTargetTriple(tm->getTargetTriple().str());
     mlib->setDataLayout(tm->createDataLayout());
+    for (llvm::Function &f : mlib->functions()) {
+      f.addFnAttr(llvm::Attribute::AlwaysInline);
+    }
     cg->AddLinkModule(std::move(mlib));
   }