[Clang][CodeGen] __builtin_allocas should care about address spaces

`alloca` instructions always return pointers to the `alloca` address space. This composes poorly with most HLLs which are address space agnostic and thus have all pointers point to generic/default. Static `alloca`s were already handled on the AST level, however dynamic `alloca`s were not, which would lead to subtly incorrect IR. This patch addresses that by inserting an address space cast iff the `alloca` address space is different from the default / expected. Reviewed By: rjmccall, arsenm Differential Revision: https://reviews.llvm.org/D156539

Author: AlexVlx

clang/lib/CodeGen/CGBuiltin.cpp

@@ -3517,6 +3517,12 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
  return RValue::get(Result);
  }
 
+ // An alloca will always return a pointer to the alloca (stack) address
+ // space. This address space need not be the same as the AST / Language
+ // default (e.g. in C / C++ auto vars are in the generic address space). At
+ // the AST level this is handled within CreateTempAlloca et al., but for the
+ // builtin / dynamic alloca we have to handle it here. We use an explicit cast
+ // instead of passing an AS to CreateAlloca so as to not inhibit optimisation.
  case Builtin::BIalloca:
  case Builtin::BI_alloca:
  case Builtin::BI__builtin_alloca_uninitialized:
@@ -3532,6 +3538,13 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
  AI->setAlignment(SuitableAlignmentInBytes);
  if (BuiltinID != Builtin::BI__builtin_alloca_uninitialized)
  initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes);
+ LangAS AAS = getASTAllocaAddressSpace();
+ LangAS EAS = E->getType()->getPointeeType().getAddressSpace();
+ if (AAS != EAS) {
+ llvm::Type *Ty = CGM.getTypes().ConvertType(E->getType());
+ return RValue::get(getTargetHooks().performAddrSpaceCast(*this, AI, AAS,
+ EAS, Ty));
+ }
  return RValue::get(AI);
  }
 
@@ -3547,6 +3560,13 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
  AI->setAlignment(AlignmentInBytes);
  if (BuiltinID != Builtin::BI__builtin_alloca_with_align_uninitialized)
  initializeAlloca(*this, AI, Size, AlignmentInBytes);
+ LangAS AAS = getASTAllocaAddressSpace();
+ LangAS EAS = E->getType()->getPointeeType().getAddressSpace();
+ if (AAS != EAS) {
+ llvm::Type *Ty = CGM.getTypes().ConvertType(E->getType());
+ return RValue::get(getTargetHooks().performAddrSpaceCast(*this, AI, AAS,
+ EAS, Ty));
+ }
  return RValue::get(AI);
  }
 

clang/test/CodeGen/dynamic-alloca-with-address-space.c

@@ -0,0 +1,41 @@
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -emit-llvm %s -o - \
+// RUN: | FileCheck %s --check-prefix=CHECK
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -DOCL12 -x cl -std=cl1.2 \
+// RUN: -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK-CL12
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -x cl -std=cl2.0 \
+// RUN: -emit-llvm %s -o - | FileCheck %s --check-prefix=CHECK-CL20
+
+#if defined(OCL12)
+ #define CAST (char *)(unsigned long)
+#else
+ #define CAST (char *)
+#endif
+
+void allocas(unsigned long n) {
+ char *a = CAST __builtin_alloca(n);
+ char *uninitialized_a = CAST __builtin_alloca_uninitialized(n);
+ char *aligned_a = CAST __builtin_alloca_with_align(n, 8);
+ char *aligned_uninitialized_a = CAST __builtin_alloca_with_align_uninitialized(n, 8);
+}
+
+// CHECK: @allocas(
+// CHECK: store i64 %n, ptr %n.addr.ascast, align 8
+// CHECK: %0 = load i64, ptr %n.addr.ascast, align 8
+// CHECK: %1 = alloca i8, i64 %0, align 8, addrspace(5)
+// CHECK: %2 = addrspacecast ptr addrspace(5) %1 to ptr
+// CHECK: store ptr %2, ptr %a.ascast, align 8
+// CHECK: %3 = load i64, ptr %n.addr.ascast, align 8
+// CHECK: %4 = alloca i8, i64 %3, align 8, addrspace(5)
+// CHECK: %5 = addrspacecast ptr addrspace(5) %4 to ptr
+// CHECK: store ptr %5, ptr %uninitialized_a.ascast, align 8
+// CHECK: %6 = load i64, ptr %n.addr.ascast, align 8
+// CHECK: %7 = alloca i8, i64 %6, align 1, addrspace(5)
+// CHECK: %8 = addrspacecast ptr addrspace(5) %7 to ptr
+// CHECK: store ptr %8, ptr %aligned_a.ascast, align 8
+// CHECK: %9 = load i64, ptr %n.addr.ascast, align 8
+// CHECK: %10 = alloca i8, i64 %9, align 1, addrspace(5)
+// CHECK: %11 = addrspacecast ptr addrspace(5) %10 to ptr
+// CHECK: store ptr %11, ptr %aligned_uninitialized_a.ascast, align 8
+// CHECK: ret void
+// CHECK-CL12-NOT: addrspacecast
+// CHECK-CL20-NOT: addrspacecast