[clang] Fix emitVoidPtrVAArg for non-zero default alloca address space

Indirect arguments are passed on the stack and so va_arg should use the
default alloca address space, not hard-code 0, for pointers to those.
The only in-tree target with a non-zero default alloca address space is
AMDGPU, but that does not support variadic arguments, so we cannot test
this upstream. However, downstream in CHERI LLVM (and Morello LLVM, a
further fork of that) we have targets that do both and so require this
change.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D132247

diff --git a/clang/lib/CodeGen/TargetInfo.cpp b/clang/lib/CodeGen/TargetInfo.cpp
index ff4d516..9fb5d9a 100644 (file)
--- a/clang/lib/CodeGen/TargetInfo.cpp
+++ b/clang/lib/CodeGen/TargetInfo.cpp
@@ -401,8 +401,10 @@ static Address emitVoidPtrVAArg(CodeGenFunction &CGF, Address VAListAddr,
 
   // Cast the address we've calculated to the right type.
   llvm::Type *DirectTy = CGF.ConvertTypeForMem(ValueTy), *ElementTy = DirectTy;
-  if (IsIndirect)
-    DirectTy = DirectTy->getPointerTo(0);
+  if (IsIndirect) {
+    unsigned AllocaAS = CGF.CGM.getDataLayout().getAllocaAddrSpace();
+    DirectTy = DirectTy->getPointerTo(AllocaAS);
+  }
 
   Address Addr = emitVoidPtrDirectVAArg(CGF, VAListAddr, DirectTy, DirectSize,
                                         DirectAlign, SlotSizeAndAlign,