[clang] [llvm] [SPIR-V] Implement SPV_KHR_untyped_pointers extension (PR #201233)

[Marcos Maronas via cfe-commits]

================
@@ -838,10 +838,37 @@ Register SPIRVGlobalRegistry::buildGlobalVariable(
   if (&GVBuilder.getMBB() != &EntryBB)
     GVBuilder.setInsertPt(EntryBB, EntryBB.getFirstTerminator());
 
-  auto MIB = GVBuilder.buildInstr(SPIRV::OpVariable)
+  // Pointers to opaque types stay typed even with the extension on, so emit the
+  // untyped variant only when the result is actually an untyped pointer.
+  const SPIRVSubtarget &Subtarget =
+      cast<SPIRVSubtarget>(MIRBuilder.getMF().getSubtarget());
+  const bool UseUntypedPointers =
+      BaseType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
+  const unsigned VariableOpcode =
+      UseUntypedPointers ? SPIRV::OpUntypedVariableKHR : SPIRV::OpVariable;
+
+  auto MIB = GVBuilder.buildInstr(VariableOpcode)
                  .addDef(ResVReg)
                  .addUse(getSPIRVTypeID(BaseType))
                  .addImm(static_cast<uint32_t>(Storage));
+
+  // OpUntypedVariableKHR takes an extra Data Type operand right after the
+  // storage class, holding the global's value type.
+  if (UseUntypedPointers) {
+    SPIRVTypeInst DataType = getPointeeType(BaseType);
+    if (!DataType)
+      DataType = getOrCreateSPIRVType(GV->getValueType(), GVBuilder,
+                                      SPIRV::AccessQualifier::ReadWrite,
+                                      /*EmitIR=*/false);
+    if (!DataType) {
+      // Use i8 as a last resort.
----------------
maarquitos14 wrote:

Is this because some corner cases we already know make it impossible to retrieve the real type, or for unforeseen cases? If the latter, should we emit a warning or add an assertion?

https://github.com/llvm/llvm-project/pull/201233