[SYCL][ESIMD] Fix crashes of sycl-post-link

[fineg74]

This is to have sycl-post-link to handle IR in the form

%0 = load i64, i64 addrspace(1)* getelementptr (<3 x i64>, <3 x i64> addrspace(1)* @__spirv_BuiltInGlobalInvocationId, i64 0, i64 0), align 32
  store i64 %0, i64 addrspace(1)* %_arg_DoNotOptimize, align 8

[fineg74]

AMDGPU failures
SYCL :: Assert/assert_in_kernels.cpp
SYCL :: Assert/assert_in_multiple_tus.cpp
SYCL :: Assert/assert_in_multiple_tus_one_ndebug.cpp
SYCL :: Assert/assert_in_one_kernel.cpp
SYCL :: Assert/assert_in_simultaneously_multiple_tus_one_ndebug.cpp

Are not related to the change

[kbobrovs]

I think this is too generic. You basically need to distinguish the
%0 = load i64, i64 addrspace(1)* getelementptr (<3 x i64>, <3 x i64> addrspace(1)* @__spirv_BuiltInGlobalInvocationId, i64 0, i64 0), align 32
case from others. This case above performs thread_id calculation in one instruction basically, unlike the other two variants, where load is followed by element extraction or truncation.

[AlexeySachkov]

The implementation is owned by ESIMD folks, only test is owned by Tools. Approving the test, so this can be merged as soon as implementation is approved.

However, I would recommend to move this test somewhere and perform it through opt. This way, it would be more like a unit-test for a single pass, because right now we are performing an integration test of the whole sycl-post-link to test a single pass change.

[kbobrovs]

Approving and merging due to urgency. But I agree with this:

However, I would recommend to move this test somewhere and perform it through opt. This way, it would be more like a unit-test for a single pass, because right now we are performing an integration test of the whole sycl-post-link to test a single pass change.

So, @fineg74 - could you please address this in later PRs?