Merge commit '2c498ee497837a2a8e8a0f285306d37528077ef4'

Author: whitneywhtsang

third_party/amd/backend/compiler.py

@@ -265,6 +265,11 @@ def make_llir(src, metadata, options):
         denormal_mode = "preserve-sign" if options.allow_flush_denorm else "ieee"
         fns[0].add_fn_attr("denormal-fp-math-f32", denormal_mode)
 
+        # Hint the compiler that we'd like the firmware to set the kernel arguments
+        # to user SGPRs so that the kernel does not need to s_load its arguments
+        # from memory.
+        amd.set_all_fn_arg_inreg(fns[0])
+
         if options.extern_libs:
             paths = [path for (name, path) in options.extern_libs if amd.need_extern_lib(llvm_mod, name)]
             llvm.link_extern_libs(llvm_mod, paths)

third_party/amd/python/triton_amd.cc

@@ -257,4 +257,13 @@ void init_triton_amd(py::module &&m) {
       return false;
     }
   });
+
+  m.def("set_all_fn_arg_inreg", [](llvm::Function *fn) {
+    for (llvm::Argument &arg : fn->args()) {
+      // Check for incompatible attributes.
+      if (arg.hasByRefAttr() || arg.hasNestAttr())
+        continue;
+      arg.addAttr(llvm::Attribute::InReg);
+    }
+  });
 }