Add check to early exit when OpenCL builtin is not found in the map

[asudarsa]

Several user defined/declared functions were being identified as 'possible' OpenCL builtins as they had prefixes (e.g. atomic_) similar to OpenCL builtin function names. However, they were not present in the OCLSPIRVBuiltinMap.
One of the OCLToSPIRV utility functions calls OCLSPIRVBuiltinMap::map function on such user defined functions and this results in a crash.
The correct way to do this will be to check if the OCL function name key is present in the map and exit early if not.

This PR does that.

Thanks

[asudarsa]

This PR should help us to get #2500 back by reverting #2508

Thanks

[bwlodarcz]

The fix of the issue is likely a little bigger than "atomic". The changes covers "atomic" prefix but there are other prefixes in OCLToSPIRVBase::visitCallInst in OCLToSPIRV.cpp:196 like "convert" or "VLoad" which needs to be handled with care as well. I didn't do deep dive but by quickly glimpsing the code of the function handling that cases it's imaginable to lead Translator again into bogus state by carefully crafted function name. What do you think @MrSidims, @asudarsa?

[asudarsa]

The fix of the issue is likely a little bigger than "atomic". The changes covers "atomic" prefix but there are other prefixes in OCLToSPIRVBase::visitCallInst in OCLToSPIRV.cpp:196 like "convert" or "VLoad" which needs to be handled with care as well. I didn't do deep dive but by quickly glimpsing the code of the function handling that cases it's imaginable to lead Translator again into bogus state by carefully crafted function name. What do you think @MrSidims, @asudarsa?

Hi @bwlodarcz

Thanks much for the feedback.
I am now trying several tests trying to manually create the user defined functions with OCL looking names. I will update the PR with my findings.

Sincerely

[asudarsa]

The fix of the issue is likely a little bigger than "atomic". The changes covers "atomic" prefix but there are other prefixes in OCLToSPIRVBase::visitCallInst in OCLToSPIRV.cpp:196 like "convert" or "VLoad" which needs to be handled with care as well. I didn't do deep dive but by quickly glimpsing the code of the function handling that cases it's imaginable to lead Translator again into bogus state by carefully crafted function name. What do you think @MrSidims, @asudarsa?

The transBuiltin(...) function already has the check to see if the function we are looking at has a valid OCL builtin (See

SPIRV-LLVM-Translator/lib/SPIRV/OCLToSPIRV.cpp

Line 915 in 5e53bcc

if (OCLSPIRVBuiltinMap::find(Info.UniqName, &OC)) {

).
This check was missing in transAtomicBuiltin(...) and this PR adds that check. I do not see any other site where this check is required. So, I think this PR can be submitted as is.

Thanks for all the feedback.

[bwlodarcz]

@asudarsa Unfortunately the change doesn't work/is not enough. Let's replace atomic_ with ndrange_ and we are entering again bogus state.

; Test that user defined/declared functions are not wrongly identified as OpenCL builtins.

target datalayout = "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-n8:16:32:64"
target triple = "spir64-unknown-unknown"

; Function Attrs: mustprogress norecurse nounwind
define weak_odr dso_local spir_kernel void @_ZTSN4test_ocl_builtin(ptr addrspace(4) noundef %arrayidx.i311) {
entry:
     %call.i312 = call spir_func noundef i32 @_Z28ndrange_fetch_and_sub_uint32Pii(ptr addrspace(4) noundef %arrayidx.i311, i32 noundef 1)
     %call.i313 = call spir_func noundef i32 @_Z28ndrange_fetch_and_add_uint32Pii(ptr addrspace(4) noundef %arrayidx.i311, i32 noundef 1)
     ret void
}

; Function Attrs: convergent inlinehint mustprogress norecurse nounwind
define linkonce_odr dso_local spir_func noundef i32 @_Z28ndrange_fetch_and_sub_uint32Pii(ptr addrspace(4) noundef %p, i32 noundef %x) {
entry:
  %call.i.i.i.i.i.i = tail call spir_func noundef ptr addrspace(1) @_Z41__spirv_GenericCastToPtrExplicit_ToGlobalPvi(ptr addrspace(4) noundef %p, i32 noundef 5)
  %0 = addrspacecast ptr addrspace(1) %call.i.i.i.i.i.i to ptr addrspace(4)
  call spir_func void @__itt_offload_atomic_op_start(ptr addrspace(4) %0, i32 2, i32 0)
  %call3.i.i = tail call spir_func noundef i32 @_Z18__spirv_AtomicISubPU3AS1iN5__spv5Scope4FlagENS1_19MemorySemanticsMask4FlagEi(ptr addrspace(1) noundef %call.i.i.i.i.i.i, i32 noundef 1, i32 noundef 896, i32 noundef %x)
  %1 = addrspacecast ptr addrspace(1) %call.i.i.i.i.i.i to ptr addrspace(4)
  call spir_func void @__itt_offload_atomic_op_finish(ptr addrspace(4) %1, i32 2, i32 0)
  ret i32 %call3.i.i
}

declare dso_local spir_func noundef i32 @_Z28ndrange_fetch_and_add_uint32Pii(ptr addrspace(4) noundef %p, i32 noundef %x)

declare dso_local spir_func noundef ptr addrspace(1) @_Z41__spirv_GenericCastToPtrExplicit_ToGlobalPvi(ptr addrspace(4) noundef, i32 noundef)

declare dso_local spir_func void @__itt_offload_atomic_op_start(ptr addrspace(4) noundef %object, i32 noundef %op_type, i32 noundef %mem_order)

declare dso_local spir_func noundef i32 @_Z18__spirv_AtomicISubPU3AS1iN5__spv5Scope4FlagENS1_19MemorySemanticsMask4FlagEi(ptr addrspace(1) noundef, i32 noundef, i32 noundef, i32 noundef)

declare dso_local spir_func void @__itt_offload_atomic_op_finish(ptr addrspace(4) noundef %object, i32 noundef %op_type, i32 noundef %mem_order)

!opencl.spir.version = !{!0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0, !0}
!spirv.Source = !{!1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1, !1}

!0 = !{i32 1, i32 2}
!1 = !{i32 4, i32 100000}

Standard llvm-as followed by llvm-spirv sequence to reproduce.
The test fails on assertions in debug in OCLToSPIRVBase::visitCallNDRange.

[bwlodarcz]

Described in previous comment.

[CLAassistant]

All committers have signed the CLA.