[clang] [llvm] [mlir] [AMDGPU] Verify AMDGPU required workgroup size matches flat workgroup size (PR #200989)

[Matt Arsenault via cfe-commits]

================
@@ -51,6 +51,9 @@ __attribute__((amdgpu_waves_per_eu(0, 4))) kernel void kernel_waves_per_eu_0_4()
 __attribute__((amdgpu_flat_work_group_size(64, 32))) kernel void kernel_flat_work_group_size_64_32() {} // expected-error {{'amdgpu_flat_work_group_size' attribute argument is invalid: min must not be greater than max}}
 __attribute__((amdgpu_waves_per_eu(4, 2))) kernel void kernel_waves_per_eu_4_2() {} // expected-error {{'amdgpu_waves_per_eu' attribute argument is invalid: min must not be greater than max}}
 
+__attribute__((reqd_work_group_size(32, 2, 1), amdgpu_flat_work_group_size(16, 128))) kernel void kernel_reqd_work_group_size_32_2_1_flat_work_group_size_16_128() {} // expected-error {{'amdgpu_flat_work_group_size' attribute must match 'reqd_work_group_size' product}}
+__attribute__((amdgpu_flat_work_group_size(16, 128), reqd_work_group_size(32, 2, 1))) kernel void kernel_flat_work_group_size_16_128_reqd_work_group_size_32_2_1() {} // expected-error {{'amdgpu_flat_work_group_size' attribute must match 'reqd_work_group_size' product}}
+
----------------
arsenm wrote:

> Even if they are allowed, they should be ignored, since they have no defined semantics outside OpenCL.

It has obvious semantics, and the good QoI would be to respect it. And merely warn if you specify Y/Z dimensions 

https://github.com/llvm/llvm-project/pull/200989