[SYCL][Test] Add IR checks into atomic tests #2834
Conversation
Check that correct `__spirv_Atomic*` external calls are emitted by the compiler FE, and thus ensure that the headers provide a proper implementation for each function. Signed-off-by: Artem Gindinson <[email protected]>
|
This is mainly being done in scope of future atomic float extensions' support. A draft example of this is #2765. |
Signed-off-by: Artem Gindinson <[email protected]>
Signed-off-by: Artem Gindinson <[email protected]>
|
@s-kanaev @intel/llvm-reviewers-runtime Ping |
| @@ -181,12 +195,31 @@ int main() { | |||
| } | |||
|
|
|||
| constexpr int N = 32; | |||
| // CHECK-LLVM: declare dso_local spir_func i32 | |||
| // CHECK-LLVM-SAME: @_Z{{[0-9]+}}__spirv_AtomicIAdd | |||
| // CHECK-LLVM-SAME: (i32 addrspace(1)*, i32, i32, i32) | |||
| add_test<int>(q, N); | |||
There was a problem hiding this comment.
I'm curious what if the function isn't inlined? Is its inlining forced somehow?
There was a problem hiding this comment.
I believe that inlining can't take place as such. These __spirv_Atomic* functions are generated due to being included from here, however no definition is provided. These external calls get lowered at later stages (e.g. mapped onto corresponding SPIR-V operands during translation from LLVM IR - see #2765 as a fresh example).
| compare_exchange_test<unsigned long long>(q, N); | ||
| // The remaining functions use the already-declared ones on the IR level |
There was a problem hiding this comment.
Do they? It was integers and now we have floats and doubles...
There was a problem hiding this comment.
The reason for this is that float/pointer-typed functions are emulated instead of being mapped onto particular SPIR-V operands. I do not know the details of why integer compare_exchange external calls are made inside this emulation, but #2765 exemplifies what happens if the emulation is changed by an external call to a unique float-typed function.
Check that correct
__spirv_Atomic*external calls are emittedby the compiler FE, and thus ensure that the headers provide a
proper implementation for each function.
Signed-off-by: Artem Gindinson [email protected]