[clang] [llvm] [llvm][opt][Transforms][SPIR-V] Enable `InferAddressSpaces` for SPIR-V (PR #110897)

[Alex Voicu via cfe-commits]

================
@@ -91,6 +97,88 @@ SPIRVTargetMachine::SPIRVTargetMachine(const Target &T, const Triple &TT,
   setRequiresStructuredCFG(false);
 }
 
+enum AddressSpace {
+  Function = storageClassToAddressSpace(SPIRV::StorageClass::Function),
+  CrossWorkgroup =
+      storageClassToAddressSpace(SPIRV::StorageClass::CrossWorkgroup),
+  UniformConstant =
+      storageClassToAddressSpace(SPIRV::StorageClass::UniformConstant),
+  Workgroup = storageClassToAddressSpace(SPIRV::StorageClass::Workgroup),
+  Generic = storageClassToAddressSpace(SPIRV::StorageClass::Generic)
+};
+
+unsigned SPIRVTargetMachine::getAssumedAddrSpace(const Value *V) const {
+  const auto *LD = dyn_cast<LoadInst>(V);
+  if (!LD)
+    return UINT32_MAX;
+
+  // It must be a load from a pointer to Generic.
+  assert(V->getType()->isPointerTy() &&
+         V->getType()->getPointerAddressSpace() == AddressSpace::Generic);
+
+  const auto *Ptr = LD->getPointerOperand();
+  if (Ptr->getType()->getPointerAddressSpace() != AddressSpace::UniformConstant)
+    return UINT32_MAX;
+  // For a loaded from a pointer to UniformConstant, we can infer CrossWorkgroup
+  // storage, as this could only have been legally initialised with a
+  // CrossWorkgroup (aka device) constant pointer.
+  return AddressSpace::CrossWorkgroup;
+}
+
+std::pair<const Value *, unsigned>
+SPIRVTargetMachine::getPredicatedAddrSpace(const Value *V) const {
+  using namespace PatternMatch;
+
+  if (auto *II = dyn_cast<IntrinsicInst>(V)) {
+    switch (II->getIntrinsicID()) {
+    case Intrinsic::amdgcn_is_shared:
+      return std::pair(II->getArgOperand(0), AddressSpace::Workgroup);
+    case Intrinsic::amdgcn_is_private:
+      return std::pair(II->getArgOperand(0), AddressSpace::Function);
+    default:
+      break;
+    }
+    return std::pair(nullptr, UINT32_MAX);
+  }
+  // Check the global pointer predication based on
+  // (!is_share(p) && !is_private(p)). Note that logic 'and' is commutative and
+  // the order of 'is_shared' and 'is_private' is not significant.
+  Value *Ptr;
+  if (getTargetTriple().getVendor() == Triple::VendorType::AMD &&
+      match(
+          const_cast<Value *>(V),
+          m_c_And(m_Not(m_Intrinsic<Intrinsic::amdgcn_is_shared>(m_Value(Ptr))),
+                m_Not(m_Intrinsic<Intrinsic::amdgcn_is_private>(
+                    m_Deferred(Ptr))))))
----------------
AlexVlx wrote:

Neither the BE nor the Translator handle that at the moment, and I suspect it's meant for implementing some specific bit of OpenCL (SYCL?) functionality. We use the non-explicit flavours, and those don't return null (and are diagnosed as illegal if they are illegal per spec). This is probably a good way of implementing the predicates / handling this, so thank you for it. Having said that, I reiterate that we have AMDGCN flavoured SPIR-V where the actual AMDGCN predicates would manifest / make sense.

https://github.com/llvm/llvm-project/pull/110897