[SYCL] Fix hier_par test after pulldown from llvm #1205
Conversation
Signed-off-by: Artur Gainullin <artur.gainullin@intel.com>
againull
force-pushed
the
hier_par_test_fix
branch
from
41e8a42 to
ceb741a
Compare
| // CHECK: %[[TMP:[A-Za-z_0-9]+]] = bitcast %[[GROUP_CLASS]] addrspace(3)* @[[SHADOW]] to i8 addrspace(3)* | ||
| // CHECK: %[[OBJ:[A-Za-z_0-9]+]] = bitcast %[[GROUP_CLASS]]* %[[GROUP_OBJ]] to i8* | ||
| // CHECK: call void @llvm.memcpy.p3i8.p0i8.i64(i8 addrspace(3)* align [[ALIGN]] %[[TMP]], {{[^,]+}} %[[OBJ]], {{[^)]+}}) | ||
| // CHECK: @[[SHADOW:[a-zA-Z0-9]+]] = internal unnamed_addr addrspace(3) global %[[GROUP_CLASS:"[^"]+"]] undef, align [[ALIGN:[0-9]+]] |
There was a problem hiding this comment.
Is LLVM transformation always uses addrspace(3) for "local" memory? I think it's supposed to be target depended.
If so, I think we should explicitly specify the target device in the RUN command to make sure that "local" memory is represented as addrspace(3).
Why do we care about align attribute?
| // CHECK: @[[SHADOW:[a-zA-Z0-9]+]] = internal unnamed_addr addrspace(3) global %[[GROUP_CLASS:"[^"]+"]] undef, align [[ALIGN:[0-9]+]] | |
| // CHECK: @[[SHADOW:[a-zA-Z0-9]+]] = internal unnamed_addr addrspace(3) global %[[GROUP_CLASS:"[^"]+"]] undef |
There was a problem hiding this comment.
Good concern, thx, I think I can create different check labels for different targets to address this.
There was a problem hiding this comment.
I'm fine if you add a TODO and we address this later.
| // CHECK: define {{.*}} spir_func void @{{"[^"]+"}}({{[^,]+}}, %[[GROUP_CLASS]]* byval(%[[GROUP_CLASS]]) align {{[0-9]+}} %[[GROUP_OBJ:[A-Za-z_0-9]+]]) {{.*}}!work_group_scope{{.*}} { | ||
| // CHECK-NOT: {{^[ \t]*define}} | ||
| // CHECK: %[[OBJ:[A-Za-z_0-9]+]] = bitcast %[[GROUP_CLASS]]* %[[GROUP_OBJ]] to i8* | ||
| // CHECK: call void @llvm.memcpy.p3i8.p0i8.i64(i8 addrspace(3)* align [[ALIGN]] getelementptr inbounds (%[[GROUP_CLASS]], %[[GROUP_CLASS]] addrspace(3)* @[[SHADOW]], i32 0, i32 0), {{[^,]+}} %[[OBJ]], {{[^)]+}}) |
There was a problem hiding this comment.
What are last two checks for?
There was a problem hiding this comment.
The last two checks are needed to ensure that group object which is passed as byval to the 'parallel for work group' is copied to 'shadow' variable which is a global variable in local address space.
There was a problem hiding this comment.
I've shared with you my private experiments with enabling standard LLVM transformation passes before translation to SPIR-V format. IIRC, these copies from private to local memory are optimized away. The same might happen in the back-end compilers.
I suggest checking if it's the case. If so, the optimized code is not valid SYCL/OpenCL program.
|
Replaced by #1306. |
Signed-off-by: Artur Gainullin artur.gainullin@intel.com