Enable inlining for dlp_execute_kernel with LTO

Add always_inline to dlp_execute_kernel to ensure it is inlined when
building with LTO. With LLVM 19, this attribute only takes effect under
LTO; in non-LTO builds, inlining is not guaranteed. This improves
performance for tiny shapes.

Author: js00070

src/frame/bindings/c_wrappers/capi_kernel_frame_wrappers.cc

@@ -239,6 +239,15 @@ dlp_init_and_get_kernel_hndl(kernel_datatype_t  k_dtype,
 // Experimentally derived alignment, needs further analysis but gives
 // consistent good performance on zen5 machines.
 [[gnu::aligned(512)]]
+// Force inlining of dlp_execute_kernel to ensure optimal performance, especially when
+// building with Link Time Optimization (LTO). Without the always_inline attribute,
+// some compilers may not inline this function even with LTO enabled, which can lead
+// to suboptimal performance in tiny shape scenarios. Explicitly marking this
+// function as always_inline guarantees that the optimizer can inline it as intended
+// when LTO is enabled.
+// Note: With LLVM 19, this attribute has no effect unless LTO is enabled; in non-LTO
+// builds, the compiler may still choose not to inline this function.
+__attribute__((always_inline))
 void
 dlp_execute_kernel(dlp_kernel_hndl_t   kernel_hndl,
                    md_t                m,