[llvm-branch-commits] [clang] [AMDGPU][Clang] add __builtin_amdgcn_av_(load|store)_b128 (PR #199176)

[Shilei Tian via llvm-branch-commits]

================
@@ -0,0 +1,26 @@
+// We test loads and stores separately because clang only seems to exit after
+// the first 'target feature' error.
+
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx602 -DTEST_LOAD  -S -verify -o - %s
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx705 -DTEST_LOAD  -S -verify -o - %s
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx810 -DTEST_LOAD  -S -verify -o - %s
+
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx602 -DTEST_STORE -S -verify -o - %s
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx705 -DTEST_STORE -S -verify -o - %s
+// RUN: %clang_cc1 -cl-std=CL2.0 -triple amdgcn-unknown-unknown -target-cpu gfx810 -DTEST_STORE -S -verify -o - %s
+// REQUIRES: amdgpu-registered-target
+
+typedef __attribute__((__vector_size__(4 * sizeof(unsigned int)))) unsigned int v4u32;
+typedef v4u32 __global *global_ptr_to_v4u32;
+
+#ifdef TEST_LOAD
----------------
shiltian wrote:

Do we really need to use macro here? There are definitely some test cases that emit errors from different functions but w/o using macro.

https://github.com/llvm/llvm-project/pull/199176