[clang] [CIR] Upstream zero init for global variables (PR #133100)

Amr Hesham via cfe-commits cfe-commits at lists.llvm.org
Thu Mar 27 06:51:13 PDT 2025


================
@@ -67,6 +67,40 @@ class CIRBaseBuilderTy : public mlir::OpBuilder {
     return create<cir::ConstantOp>(loc, attr.getType(), attr);
   }
 
+  mlir::TypedAttr getConstNullPtrAttr(mlir::Type t) {
+    assert(mlir::isa<cir::PointerType>(t) && "expected cir.ptr");
+    return getConstPtrAttr(t, 0);
+  }
+
+  mlir::TypedAttr getZeroAttr(mlir::Type t) {
+    return cir::ZeroAttr::get(getContext(), t);
+  }
+
+  mlir::TypedAttr getZeroInitAttr(mlir::Type ty) {
+    if (mlir::isa<cir::IntType>(ty))
+      return cir::IntAttr::get(ty, 0);
+    if (auto fltType = mlir::dyn_cast<cir::SingleType>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto fltType = mlir::dyn_cast<cir::DoubleType>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto fltType = mlir::dyn_cast<cir::LongDoubleType>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto fltType = mlir::dyn_cast<cir::FP16Type>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto fltType = mlir::dyn_cast<cir::FP128Type>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto fltType = mlir::dyn_cast<cir::BF16Type>(ty))
+      return cir::FPAttr::getZero(fltType);
+    if (auto arrTy = mlir::dyn_cast<cir::ArrayType>(ty))
+      return getZeroAttr(arrTy);
+    if (auto ptrTy = mlir::dyn_cast<cir::PointerType>(ty))
+      return getConstNullPtrAttr(ptrTy);
+    if (mlir::isa<cir::BoolType>(ty)) {
+      return getCIRBoolAttr(false);
+    }
+    llvm_unreachable("Zero initializer for given type is NYI");
+  }
----------------
AmrDeveloper wrote:

> This was probably done in response to my suggestion that handling the zero-initialization case for arrays would be a way to make the other PR smaller. I didn't realize this piece of it was so small.

I was thinking of extracting the zero attr part into another PR, and if it can be merged first, then I will update the other patch. but I think now it's doesn't matter which one will be merged first it's easy to update the other one 

https://github.com/llvm/llvm-project/pull/133100


More information about the cfe-commits mailing list