[PR40778] Generate address space conversion when binding reference to a temporary value in different address space

[Anastasia]

Summary

This is fixing one of the issues reported in the bug:
https://bugs.llvm.org/show_bug.cgi?id=40778

[Anastasia]

lib/CodeGen/CGCall.cpp

4067 ↗

(On Diff #188206)

We have started using performAddrSpaceCast for those but, however, I was very confused about how to get address space of the Clang types correctly here.

I was using this function in a couple of places before but I must admit I am still a bit confused about the purpose of it. Mainly I was wondering if we could drop LangAS parameters from it? It would simplify its use a lot in multiple places. As far as I can see they are not used in the function at the moment. I am not sure if they are reserved for some development in the future though?

Altogether I quite like using IRBuilder directly because in combination with target address space map it allows to do what's needed here and keep consistent with the rest. Perhaps, I am missing some background info. I have tried to dig out a bit but no much success.

[rjmccall]

lib/CodeGen/CGCall.cpp

4067 ↗

(On Diff #188206)

We want to allow language address spaces to be arbitrarily mapped to IR address spaces during lowering, which might include e.g. changing the null value. That's why we pass language address spaces down.

You should be getting the language address space from the appropriate type. I wouldn't expect that to include address-space conversions at this point, though; is there maybe a missing implicit conversions somewhere?

[Anastasia]

lib/CodeGen/CGCall.cpp

4067 ↗

(On Diff #188206)

We want to allow language address spaces to be arbitrarily mapped to IR address spaces during lowering, which might include e.g. changing the null value. That's why we pass language address spaces down.

Would this mean we can map AST address spaces into different IR ones from what is the the target address space map? Ok right, null pointer value can't be handled by the target address space map at all. I am a bit confused though why we are not using these LangAS at the moment. I guess we just don't have upstream targets requiring that...

is there maybe a missing implicit conversions somewhere?

Ok, the AST dump doesn't seem right for the test case I included. MaterializeTemporaryExpr is in generic addr space. I guess it should be private... and then converted to generic.

`-FunctionDecl 0x712608 <line:5:1, line:11:1> line:5:6 foo 'void ()'
  `-CompoundStmt 0x712810 <col:11, line:11:1>
    `-ExprWithCleanups 0x7127f8 <line:10:3, col:8> 'int'
      `-CallExpr 0x7127a0 <col:3, col:8> 'int'
        |-ImplicitCastExpr 0x712788 <col:3> 'int (*)(const __generic unsigned int &)' <FunctionToPointerDecay>
        | `-DeclRefExpr 0x712710 <col:3> 'int (const __generic unsigned int &)' lvalue Function 0x7124d0 'bar' 'int (const __generic unsigned int &)'
        `-MaterializeTemporaryExpr 0x7127e0 <col:7> 'const __generic unsigned int' lvalue
          `-ImplicitCastExpr 0x7127c8 <col:7> 'const __generic unsigned int' <IntegralCast>
            `-IntegerLiteral 0x7126f0 <col:7> 'int' 1

I will fix it!

[Anastasia]

• Implement the fix correctly by added an extra address space conversion step after binding the reference

[rjmccall]

This test file will probably grow over time, so please add a CHECK-LABEL line to the test case to make sure you're checking the body of foo(). Also, you'll probably need to allow the name %ref.tmp to be stripped for this test to pass in release builds; please go ahead and test a more complete pattern, including the alloca and the store of 1 to it. Otherwise LGTM.

[ebevhan]

lib/Sema/SemaInit.cpp

4806–4808

Sorry for the really late comment on this, but shouldn't this be:

if (RefRelationship == Sema::Ref_Related &&
    ((T1CVRQuals | T2CVRQuals) != T1CVRQuals ||
     !T1Quals.isAddressSpaceSupersetOf(T2Quals))) {

Currently, this fails on AS qualification regardless of ref-compatibility.

[Anastasia]

lib/Sema/SemaInit.cpp

4806–4808

Sorry for noticing this late. I agree this doesn't look right... I will create a patch to fix this.

include/clang/AST/Type.h

Qualifiers withoutObjCGCAttr() const {
     qs.removeObjCGCAttr();
     return qs;
   }
   Qualifiers withoutObjCLifetime() const {
     Qualifiers qs = *this;
     qs.removeObjCLifetime();
     return qs;
   }
+  Qualifiers withoutAddressSpace() const {
+    Qualifiers qs = *this;
+    qs.removeAddressSpace();
+    return qs;
+  }
 
   bool hasObjCLifetime() const { return Mask & LifetimeMask; }
   ObjCLifetime getObjCLifetime() const {
     return ObjCLifetime((Mask & LifetimeMask) >> LifetimeShift);
   }
   void setObjCLifetime(ObjCLifetime type) {
     Mask = (Mask & ~LifetimeMask) | (type << LifetimeShift);
   }

lib/Sema/SemaInit.cpp

if (T2->isRecordType()) {
     return;
   }
 
   //      - Otherwise, a temporary of type "cv1 T1" is created and initialized
   //        from the initializer expression using the rules for a non-reference
   //        copy-initialization (8.5). The reference is then bound to the
   //        temporary. [...]
 
-  InitializedEntity TempEntity = InitializedEntity::InitializeTemporary(cv1T1);
+  // Ignore address space of reference type at this point and perform address
+  // space conversion after the reference binding step.
+  QualType cv1T1IgnoreAS =
+      T1Quals.hasAddressSpace()
+          ? S.Context.getQualifiedType(T1, T1Quals.withoutAddressSpace())
+          : cv1T1;
+
+  InitializedEntity TempEntity =
+      InitializedEntity::InitializeTemporary(cv1T1IgnoreAS);
 
   // FIXME: Why do we use an implicit conversion here rather than trying
   // copy-initialization?
   ImplicitConversionSequence ICS
     = S.TryImplicitConversion(Initializer, TempEntity.getType(),
                               /*SuppressUserConversions=*/false,
                               /*AllowExplicit=*/false,
                               /*FIXME:InOverloadResolution=*/false,
if (ICS.isBad()) {
     Sequence.AddConversionSequenceStep(ICS, TempEntity.getType());
   }
 
   //        [...] If T1 is reference-related to T2, cv1 must be the
   //        same cv-qualification as, or greater cv-qualification
   //        than, cv2; otherwise, the program is ill-formed.
   unsigned T1CVRQuals = T1Quals.getCVRQualifiers();
   unsigned T2CVRQuals = T2Quals.getCVRQualifiers();
-  if (RefRelationship == Sema::Ref_Related &&
-      (T1CVRQuals | T2CVRQuals) != T1CVRQuals) {
+  if ((RefRelationship == Sema::Ref_Related &&
+       (T1CVRQuals | T2CVRQuals) != T1CVRQuals) ||
+      !T1Quals.isAddressSpaceSupersetOf(T2Quals)) {
     Sequence.SetFailed(InitializationSequence::FK_ReferenceInitDropsQualifiers);
     return;
   }
 
   //   [...] If T1 is reference-related to T2 and the reference is an rvalue
   //   reference, the initializer expression shall not be an lvalue.
   if (RefRelationship >= Sema::Ref_Related && !isLValueRef &&
       InitCategory.isLValue()) {
     Sequence.SetFailed(
                     InitializationSequence::FK_RValueReferenceBindingToLValue);
     return;
   }
 
-  Sequence.AddReferenceBindingStep(cv1T1, /*bindingTemporary=*/true);
+  Sequence.AddReferenceBindingStep(cv1T1IgnoreAS, /*bindingTemporary=*/true);
+
+  if (T1Quals.hasAddressSpace())
+    Sequence.AddQualificationConversionStep(cv1T1, isLValueRef ? VK_LValue
+                                                               : VK_XValue);
 }
 
 /// Attempt character array initialization from a string literal
 /// (C++ [dcl.init.string], C99 6.7.8).
 static void TryStringLiteralInitialization(Sema &S,
                                            const InitializedEntity &Entity,
                                            const InitializationKind &Kind,
                                            Expr *Initializer,

test/CodeGenOpenCLCXX/addrspace-references.cl

+//RUN: %clang_cc1 %s -cl-std=c++ -triple spir -emit-llvm -o - | FileCheck %s
+
+int bar(const unsigned int &i);
+// CHECK-LABEL: define spir_func void @_Z3foov() 
+void foo() {
+  // The generic addr space reference parameter object will be bound
+  // to a temporary value allocated in private addr space. We need an
+  // addrspacecast before passing the value to the function.
+  // CHECK: [[REF:%.*]] = alloca i32
+  // CHECK: store i32 1, i32* [[REF]]
+  // CHECK: [[REG:%[0-9]+]] = addrspacecast i32* [[REF]] to i32 addrspace(4)*
+  // CHECK: call spir_func i32 @_Z3barRU3AS4Kj(i32 addrspace(4)* nonnull dereferenceable(4) [[REG]])
+  bar(1);
+}