[SYCL][ESIMD][E2E] Fix noinline_args tests after pulldown (intel#11327)

After the recent pulldown, clang generates llvm.lifetime calls for the
`add` function arguments in different locations causing us to hit a
known optimizable case in ESIMDOptimizeVecArgCallConv so we start
passing the arguments by value. This ends up causing different stack
size, and these tests are very sensitive to stack size.

Add a noop inline assembly statement to prevent the optimization from
kicking in so we preserve by-pointer argument passing in this case.

Closes: intel#11036

Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

sycl/test-e2e/ESIMD/noinline_args_size/Inputs/noinline_args_size_common.hpp

@@ -31,6 +31,12 @@ class KernelID;
 
 template <typename TA, typename TB, typename TC>
 ESIMD_NOINLINE TC add(TA A, TB B) {
+#ifdef __SYCL_DEVICE_ONLY__
+  // Noop inline asm to prevent arguments being
+  // optimized to be pass by value as this
+  // test is very sensitive to stack size.
+  __asm__("" : : "r"(A.data()), "r"(B.data()));
+#endif
   return (TC)A + (TC)B;
 }
 

sycl/test-e2e/ESIMD/noinline_args_size/noinline_args_char_int_size_64.cpp

@@ -5,7 +5,6 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// XFAIL: gpu && linux
 // RUN: %{build} -I%S/.. -o %t.out
 // RUN: env IGC_FunctionControl=3 IGC_ForceInlineStackCallWithImplArg=1 %{run} %t.out