diff --git a/test/Pipeline/i64-memory-alignment.forth b/test/Pipeline/i64-memory-alignment.forth
index e36113e..ab95703 100644
--- a/test/Pipeline/i64-memory-alignment.forth
+++ b/test/Pipeline/i64-memory-alignment.forth
@@ -1,8 +1,8 @@
 \ RUN: %warpforthc %s -o %t.ptx
 \ RUN: %FileCheck %s --implicit-check-not='{{(ld|st)[.a-z]*\.[bsuf]32}}' < %t.ptx
 
-\ i64 global accesses and the stack slots in word functions are naturally
-\ aligned, so each is a single 64-bit memory operation.
+\ i64 global accesses are naturally aligned, so each is a single 64-bit memory
+\ operation.
 
 \! kernel main
 \! param A i64[256]
@@ -18,7 +18,3 @@ GLOBAL-ID CELLS A + !
 \ CHECK: ld{{(\.global)?}}.{{[bsuf]}}64 {{%[a-z]+[0-9]+}}, [
 \ CHECK: ld{{(\.global)?}}.{{[bsuf]}}64 {{%[a-z]+[0-9]+}}, [
 \ CHECK: st{{(\.global)?}}.{{[bsuf]}}64 [
-
-\ CHECK-LABEL: .func {{.*}}SUM3(
-\ CHECK: ld{{(\.global)?}}.{{[bsuf]}}64 {{%[a-z]+[0-9]+}}, [
-\ CHECK: st{{(\.global)?}}.{{[bsuf]}}64 [
diff --git a/test/Pipeline/inlined-words.forth b/test/Pipeline/inlined-words.forth
new file mode 100644
index 0000000..239bfe1
--- /dev/null
+++ b/test/Pipeline/inlined-words.forth
@@ -0,0 +1,13 @@
+\ RUN: %warpforthc %s -o %t.ptx
+\ RUN: %FileCheck %s < %t.ptx
+
+\ Words inlined into the kernel must not survive as device functions.
+\ CHECK-NOT: .func
+\ CHECK: .entry k(
+\ CHECK-NOT: .func
+
+\! kernel k
+\! param a i64[256]
+: sq DUP * ;
+: sumsq 0 SWAP 0 DO I sq + LOOP ;
+GLOBAL-ID sumsq GLOBAL-ID CELLS a + !
diff --git a/tools/warpforthc/CMakeLists.txt b/tools/warpforthc/CMakeLists.txt
index 9b8c8fc..08a225a 100644
--- a/tools/warpforthc/CMakeLists.txt
+++ b/tools/warpforthc/CMakeLists.txt
@@ -12,6 +12,7 @@ target_link_libraries(warpforthc
   MLIRArithToLLVM
   MLIRBuiltinToLLVMIRTranslation
   MLIRControlFlowToLLVM
+  MLIRFuncInlinerExtension
   MLIRFuncToLLVM
   MLIRGPUToLLVMIRTranslation
   MLIRLLVMToLLVMIRTranslation
diff --git a/tools/warpforthc/warpforthc.cpp b/tools/warpforthc/warpforthc.cpp
index 0815d67..bd13536 100644
--- a/tools/warpforthc/warpforthc.cpp
+++ b/tools/warpforthc/warpforthc.cpp
@@ -9,8 +9,9 @@
 #include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVM.h"
 #include "mlir/Conversion/MemRefToLLVM/MemRefToLLVM.h"
 #include "mlir/Dialect/ControlFlow/IR/ControlFlow.h" // IWYU pragma: keep
-#include "mlir/Dialect/Func/IR/FuncOps.h"            // IWYU pragma: keep
-#include "mlir/Dialect/MemRef/IR/MemRef.h"           // IWYU pragma: keep
+#include "mlir/Dialect/Func/Extensions/InlinerExtension.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"  // IWYU pragma: keep
+#include "mlir/Dialect/MemRef/IR/MemRef.h" // IWYU pragma: keep
 #include "mlir/IR/AsmState.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "mlir/IR/Diagnostics.h"
@@ -75,6 +76,7 @@ int main(int argc, char **argv) {
   cf::registerConvertControlFlowToLLVMInterface(registry);
   registerConvertFuncToLLVMInterface(registry);
   registerConvertMemRefToLLVMInterface(registry);
+  func::registerInlinerExtension(registry);
 
   registerBuiltinDialectTranslation(registry);
   registerGPUDialectTranslation(registry);