Refactor handling of untyped pointers in builtins reverse translation #2924
Conversation
test/transcoding/enqueue_kernel.cl
Outdated
| @@ -72,6 +85,8 @@ kernel void device_side_enqueue(global int *a, global int *b, int i, char c0) { | |||
| // CHECK-LLVM: [[BlockInv2:%[0-9]+]] = addrspacecast ptr @__device_side_enqueue_block_invoke_kernel to ptr addrspace(4) | |||
| // CHECK-LLVM: call spir_func i32 @__enqueue_kernel_basic(ptr {{.*}}, i32 {{.*}}, ptr {{.*}}, ptr addrspace(4) [[BlockInv2]], ptr addrspace(4) [[Block2Ptr]]) | |||
| // CHECK-SPV-IR: call spir_func i32 @_Z21__spirv_EnqueueKernelP13__spirv_Queuei9ndrange_tiPU3AS4P19__spirv_DeviceEventS5_U13block_pointerFvvEPU3AS4cii(target("spirv.Queue") {{.*}}, i32 {{.*}}, ptr {{.*}}, i32 0, ptr addrspace(4) null, ptr addrspace(4) null, ptr @__device_side_enqueue_block_invoke_kernel, ptr addrspace(4) {{.*}}, i32 {{.*}}, i32 {{.*}}) | |||
| // _Z21__spirv_EnqueueKernelP13__spirv_Queuei9ndrange_tiPU3AS4P19__spirv_DeviceEventS5_U13block_pointerFvvEPU3AS412structtype.0ii | |||
| // Actuallly the second mangling seems correct because the pointer really points to the structure, not to char. However, SPIR-V spec says this arg should be a pointer to char. | |||
There was a problem hiding this comment.
| // Actuallly the second mangling seems correct because the pointer really points to the structure, not to char. However, SPIR-V spec says this arg should be a pointer to char. | |
| // Actually the second mangling seems correct because the pointer really points to the structure, not to char. However, SPIR-V spec says this arg should be a pointer to char. |
or was it left just for the review? :)
There was a problem hiding this comment.
You talk here about Param operand of OpEnqueueKernel, right? What is the type of it in the generated by the translator SPIR-V module?
About the mangling, I believe, that if the SPIR-V spec mandates, that Param is char *, the SPIR-V friendly LLVM IR must follow and hence the mangling should not encode the struct type.
There was a problem hiding this comment.
Yep, this comment was for review :)
In SPIR-V we have untyped pointer (which untyped variable points to struct type). With "usual" pointers we have the same pointer to struct, just casted to pointer to char.
I believe you're right, we should not encode the struct type in this special case.
This fixes the mangling in
enqueue_kernel.cltest.Without the fix we get such mangled function:
instead of the expected: