[SYCL][ESIMD][Driver] Pass -fsycl-esimd-force-stateless-mem to host (#9825)

Right now, the `-fsycl-esimd-force-stateless-mem` flag sets the
`__ESIMD_FORCE_STATELESS_MEM` macro for the device compiler only. We
have some APIs that have different arguments with
`__ESIMD_FORCE_STATELESS_MEM` vs without it, so getting the host
compiler to not error when calling one of those functions can be
frustrating for the user.

Make `-fsycl-esimd-force-stateless-mem` set
`__ESIMD_FORCE_STATELESS_MEM` for the host compiler too.

I also found we had some esimd_emulator tests that were testing
stateless mode, but the emulator never uses the device code and always
uses host code, so it wasn't even testing stateless mode. We discussed
this internally and decided to disable the tests on the emulator.

---------

Signed-off-by: Sarnie, Nick <[email protected]>

Author: sarnex

clang/lib/Driver/ToolChains/Clang.cpp

@@ -5129,10 +5129,6 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
       CmdArgs.push_back("-fsycl-allow-func-ptr");
     }
 
-    if (Args.hasFlag(options::OPT_fsycl_esimd_force_stateless_mem,
-                     options::OPT_fno_sycl_esimd_force_stateless_mem, false))
-      CmdArgs.push_back("-fsycl-esimd-force-stateless-mem");
-
     // Forward -fsycl-instrument-device-code option to cc1. This option will
     // only be used for SPIR-V-based targets.
     if (Triple.isSPIR())
@@ -5318,6 +5314,9 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
       for (auto &Macro : D.getSYCLTargetMacroArgs())
         CmdArgs.push_back(Args.MakeArgString(Macro));
     }
+    if (Args.hasFlag(options::OPT_fsycl_esimd_force_stateless_mem,
+                     options::OPT_fno_sycl_esimd_force_stateless_mem, false))
+      CmdArgs.push_back("-fsycl-esimd-force-stateless-mem");
   }
 
   if (IsOpenMPDevice) {

clang/lib/Frontend/InitPreprocessor.cpp

@@ -1309,13 +1309,13 @@ static void InitializePredefinedMacros(const TargetInfo &TI,
       Builder.defineMacro("__ENABLE_USM_ADDR_SPACE__");
       Builder.defineMacro("SYCL_DISABLE_FALLBACK_ASSERT");
     }
-
-    if (LangOpts.SYCLESIMDForceStatelessMem)
-      Builder.defineMacro("__ESIMD_FORCE_STATELESS_MEM");
   }
   if (LangOpts.SYCLUnnamedLambda)
     Builder.defineMacro("__SYCL_UNNAMED_LAMBDA__");
 
+  if (LangOpts.SYCLESIMDForceStatelessMem)
+    Builder.defineMacro("__ESIMD_FORCE_STATELESS_MEM");
+
   // OpenCL definitions.
   if (LangOpts.OpenCL) {
     InitializeOpenCLFeatureTestMacros(TI, LangOpts, Builder);

clang/test/Driver/sycl-esimd-force-stateless-mem.cpp

@@ -1,12 +1,12 @@
 
 /// Verify that the driver option is translated to corresponding options
-/// to device compilation and sycl-post-link.
+/// to host/device compilation and sycl-post-link.
 // RUN: %clang -### -fsycl -fsycl-esimd-force-stateless-mem \
 // RUN: %s 2>&1 | FileCheck -check-prefix=CHECK-PASS-TO-COMPS %s
 // CHECK-PASS-TO-COMPS: clang{{.*}} "-fsycl-esimd-force-stateless-mem"
 // CHECK-PASS-TO-COMPS: sycl-post-link{{.*}} "-lower-esimd-force-stateless-mem"
-// CHECK-PASS-TO-COMPS-NOT: clang{{.*}} "-fsycl-is-host" {{.*}}"-fsycl-esimd-force-stateless-mem"
-// CHECK-PASS-TO-COMPS-NOT: clang{{.*}} "-fsycl-esimd-force-stateless-mem" {{.*}}"-fsycl-is-host"
+// CHECK-PASS-TO-COMPS: clang{{.*}} "-fsycl-is-host" {{.*}}"-fsycl-esimd-force-stateless-mem"
+"
 
 /// Verify that stateless memory accesses mapping is not enforced by default
 // RUN: %clang -### -fsycl %s 2>&1 | FileCheck -check-prefix=CHECK-DEFAULT %s

sycl/test-e2e/ESIMD/acc_gather_scatter_rgba_stateless.cpp

@@ -5,6 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+// UNSUPPORTED: esimd_emulator
 // Use -O2 to avoid huge stack usage under -O0.
 // RUN: %{build} -O2 -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out

sycl/test-e2e/ESIMD/accessor_gather_scatter_stateless.cpp

@@ -5,6 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+// UNSUPPORTED: esimd_emulator
 // Use -O2 to avoid huge stack usage under -O0.
 // RUN: %{build} -O2 -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out

sycl/test-e2e/ESIMD/accessor_load_store_stateless.cpp

@@ -10,6 +10,7 @@
 // intrinsics when stateless memory accesses are enforced, i.e. accessor
 // based accesses are automatically converted to stateless accesses.
 
+// UNSUPPORTED: esimd_emulator
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 

sycl/test-e2e/ESIMD/accessor_stateless.cpp

@@ -5,6 +5,7 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+// UNSUPPORTED: esimd_emulator
 // Use -O2 to avoid huge stack usage under -O0.
 // RUN: %{build} -O2 -fsycl-esimd-force-stateless-mem -D_CRT_SECURE_NO_WARNINGS=1 -o %t.out
 // RUN: %{run} %t.out

sycl/test-e2e/ESIMD/lsc/lsc_predicate_stateless.cpp

@@ -5,7 +5,8 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 

sycl/test-e2e/ESIMD/lsc/lsc_surf_load_u32_stateless.cpp

@@ -5,8 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 
-#include "lsc_surf_load_u32.cpp"
+#include "lsc_surf_load_u32.cpp"

sycl/test-e2e/ESIMD/lsc/lsc_surf_load_u64_stateless.cpp

@@ -5,8 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 
-#include "lsc_surf_load_u64.cpp"
+#include "lsc_surf_load_u64.cpp"

sycl/test-e2e/ESIMD/lsc/lsc_surf_load_u8_u16_stateless.cpp

@@ -5,8 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 
-#include "lsc_surf_load_u8_u16.cpp"
+#include "lsc_surf_load_u8_u16.cpp"

sycl/test-e2e/ESIMD/lsc/lsc_surf_store_u32_stateless.cpp

@@ -5,8 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 
-#include "lsc_surf_store_u32.cpp"
+#include "lsc_surf_store_u32.cpp"

sycl/test-e2e/ESIMD/lsc/lsc_surf_store_u64_stateless.cpp

@@ -5,8 +5,9 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
-// REQUIRES: gpu-intel-pvc || esimd_emulator
+// UNSUPPORTED: esimd_emulator
+// REQUIRES: gpu-intel-pvc
 // RUN: %{build} -fsycl-esimd-force-stateless-mem -o %t.out
 // RUN: %{run} %t.out
 
-#include "lsc_surf_store_u64.cpp"
+#include "lsc_surf_store_u64.cpp"