[clang] [CIR] Pointer and vptr width from a CIR-native data-layout entry (PR #204185)
Akshay K via cfe-commits
cfe-commits at lists.llvm.org
Tue Jun 16 08:48:22 PDT 2026
https://github.com/kumarak updated https://github.com/llvm/llvm-project/pull/204185
>From a8e24830643e29745fb6efc97c9e5cb1148c7c4a Mon Sep 17 00:00:00 2001
From: AkshayK <iit.akshay at gmail.com>
Date: Tue, 16 Jun 2026 11:14:06 -0400
Subject: [PATCH] [CIR] Drive pointer and vptr width from a CIR-native
data-layout entry
PointerType and VPtrType hardcoded size/alignment to 64/8, aborting record
layout on 32-bit-pointer targets such as nvptx/spirv32. Attach a CIR-native
cir.ptr data-layout entry at module setup and read the pointer width from it,
so the CIR type system needs no LLVM-dialect dependency; the entry is stripped
during CIR->LLVM lowering. 64-bit targets are unchanged.
---
clang/lib/CIR/CodeGen/CIRGenerator.cpp | 25 ++++++++-
clang/lib/CIR/Dialect/IR/CIRTypes.cpp | 56 +++++++++++++++++--
.../CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp | 17 ++++++
.../test/CIR/CodeGen/pointer-width-32bit.cpp | 44 +++++++++++++++
4 files changed, 135 insertions(+), 7 deletions(-)
create mode 100644 clang/test/CIR/CodeGen/pointer-width-32bit.cpp
diff --git a/clang/lib/CIR/CodeGen/CIRGenerator.cpp b/clang/lib/CIR/CodeGen/CIRGenerator.cpp
index d4fcbb6e42f3e..61efaebad9b82 100644
--- a/clang/lib/CIR/CodeGen/CIRGenerator.cpp
+++ b/clang/lib/CIR/CodeGen/CIRGenerator.cpp
@@ -21,6 +21,7 @@
#include "clang/AST/DeclGroup.h"
#include "clang/CIR/CIRGenerator.h"
#include "clang/CIR/InitAllDialects.h"
+#include "clang/CIR/MissingFeatures.h"
#include "llvm/IR/DataLayout.h"
using namespace cir;
@@ -42,7 +43,29 @@ static void setMLIRDataLayout(mlir::ModuleOp &mod, const llvm::DataLayout &dl) {
mlir::MLIRContext *mlirContext = mod.getContext();
mlir::DataLayoutSpecInterface dlSpec =
mlir::translateDataLayout(dl, mlirContext);
- mod->setAttr(mlir::DLTIDialect::kDataLayoutAttrName, dlSpec);
+
+ // Add a CIR-native pointer data-layout entry so cir.ptr / cir.vptr size and
+ // alignment are driven by the data layout rather than hardcoded.
+ // The value stores {size-in-bits, abi-align-in-bits} keyed on cir.ptr.
+ //
+ // TODO(cir): Only the default address space is recorded and
+ // address-space-dependent pointer sizes are not modeled yet. Emit
+ // per-address-space entries.
+ assert(!cir::MissingFeatures::dataLayoutPtrHandlingBasedOnLangAS());
+ constexpr unsigned kBitsInByte = 8;
+ unsigned ptrSizeBits = dl.getPointerSizeInBits(/*AS=*/0);
+ unsigned ptrAlignBits =
+ dl.getPointerABIAlignment(/*AS=*/0).value() * kBitsInByte;
+ auto ptrKey = cir::PointerType::get(cir::VoidType::get(mlirContext));
+ auto ptrVal = mlir::DenseI32ArrayAttr::get(
+ mlirContext,
+ {static_cast<int32_t>(ptrSizeBits), static_cast<int32_t>(ptrAlignBits)});
+ llvm::SmallVector<mlir::DataLayoutEntryInterface> entries(
+ dlSpec.getEntries().begin(), dlSpec.getEntries().end());
+ entries.push_back(mlir::DataLayoutEntryAttr::get(ptrKey, ptrVal));
+
+ mod->setAttr(mlir::DLTIDialect::kDataLayoutAttrName,
+ mlir::DataLayoutSpecAttr::get(mlirContext, entries));
}
void CIRGenerator::Initialize(ASTContext &astContext) {
diff --git a/clang/lib/CIR/Dialect/IR/CIRTypes.cpp b/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
index e2e754479c654..041d5e1881596 100644
--- a/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
+++ b/clang/lib/CIR/Dialect/IR/CIRTypes.cpp
@@ -580,20 +580,62 @@ void RecordType::removeABIConversionNamePrefix() {
// Data Layout information for types
//===----------------------------------------------------------------------===//
+// A CIR-native pointer data-layout entry stores {size-in-bits,
+// abi-align-in-bits} as a dense i32 array keyed on a cir.ptr type (see
+// setMLIRDataLayout in CIRGenerator).
+namespace {
+constexpr static uint64_t kBitsInByte = 8;
+
+// Defaults used only when the module carries no cir.ptr data-layout entry
+// (e.g. CIR parsed from text without a data layout). These mirror the MLIR LLVM
+// dialect's pointer defaults.
+constexpr static uint64_t kDefaultPointerSizeBits = 64;
+constexpr static uint64_t kDefaultPointerAlignment = 8;
+
+enum class CIRPtrDLPos { Size = 0, AbiAlign = 1 };
+
+// Returns the requested field of the cir.ptr data-layout entry.
+std::optional<uint64_t> getPointerSpecValue(mlir::DataLayoutEntryListRef params,
+ CIRPtrDLPos pos) {
+ for (mlir::DataLayoutEntryInterface entry : params) {
+ if (!entry.isTypeEntry())
+ continue;
+ auto spec = mlir::dyn_cast<mlir::DenseI32ArrayAttr>(entry.getValue());
+ assert(spec && spec.size() == 2 &&
+ "malformed cir.ptr data layout entry: expected a pair of i32 "
+ "{size-in-bits, abi-align-in-bits}");
+ return static_cast<uint64_t>(spec[static_cast<int>(pos)]);
+ }
+ return std::nullopt;
+}
+} // namespace
+
llvm::TypeSize
PointerType::getTypeSizeInBits(const ::mlir::DataLayout &dataLayout,
::mlir::DataLayoutEntryListRef params) const {
+ // The pointer width comes from the CIR-native data-layout entry keyed on
+ // cir.ptr, which records the width for the default address space; fall back
+ // to 64 bits if the module carries no such entry.
// FIXME: improve this in face of address spaces
assert(!cir::MissingFeatures::dataLayoutPtrHandlingBasedOnLangAS());
- return llvm::TypeSize::getFixed(64);
+ if (std::optional<uint64_t> size =
+ getPointerSpecValue(params, CIRPtrDLPos::Size))
+ return llvm::TypeSize::getFixed(*size);
+ return llvm::TypeSize::getFixed(kDefaultPointerSizeBits);
}
uint64_t
PointerType::getABIAlignment(const ::mlir::DataLayout &dataLayout,
::mlir::DataLayoutEntryListRef params) const {
+ // As with the size, the alignment is taken from the default-address-space
+ // cir.ptr data-layout entry. Address-space-dependent alignments are not yet
+ // modeled.
// FIXME: improve this in face of address spaces
assert(!cir::MissingFeatures::dataLayoutPtrHandlingBasedOnLangAS());
- return 8;
+ if (std::optional<uint64_t> align =
+ getPointerSpecValue(params, CIRPtrDLPos::AbiAlign))
+ return *align / kBitsInByte;
+ return kDefaultPointerAlignment;
}
llvm::TypeSize
@@ -1112,14 +1154,16 @@ DataMemberType::getABIAlignment(const ::mlir::DataLayout &dataLayout,
llvm::TypeSize
VPtrType::getTypeSizeInBits(const mlir::DataLayout &dataLayout,
mlir::DataLayoutEntryListRef params) const {
- // FIXME: consider size differences under different ABIs
- return llvm::TypeSize::getFixed(64);
+ // The vtable pointer is an ordinary data pointer; route the query through a
+ // cir.ptr so it picks up the same data-layout-driven width.
+ return dataLayout.getTypeSizeInBits(
+ cir::PointerType::get(cir::VoidType::get(getContext())));
}
uint64_t VPtrType::getABIAlignment(const mlir::DataLayout &dataLayout,
mlir::DataLayoutEntryListRef params) const {
- // FIXME: consider alignment differences under different ABIs
- return 8;
+ return dataLayout.getTypeABIAlignment(
+ cir::PointerType::get(cir::VoidType::get(getContext())));
}
//===----------------------------------------------------------------------===//
diff --git a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
index 1579e967885d8..dd244cc4ef641 100644
--- a/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
+++ b/clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp
@@ -3689,6 +3689,23 @@ void ConvertCIRToLLVMPass::runOnOperation() {
if (failed(applyPartialConversion(ops, target, std::move(patterns))))
signalPassFailure();
+ // The CIR-native pointer data-layout entry (keyed on cir.ptr) drives pointer
+ // widths during CIR codegen and lowering, but cir.ptr has no meaning once the
+ // module is translated to LLVM IR. Drop it so the resulting data layout only
+ // references LLVM types.
+ if (auto dlSpec = mlir::dyn_cast_or_null<mlir::DataLayoutSpecAttr>(
+ module->getAttr(mlir::DLTIDialect::kDataLayoutAttrName))) {
+ llvm::SmallVector<mlir::DataLayoutEntryInterface> kept;
+ for (mlir::DataLayoutEntryInterface entry : dlSpec.getEntries()) {
+ if (entry.isTypeEntry() &&
+ mlir::isa<cir::PointerType>(mlir::cast<mlir::Type>(entry.getKey())))
+ continue;
+ kept.push_back(entry);
+ }
+ module->setAttr(mlir::DLTIDialect::kDataLayoutAttrName,
+ mlir::DataLayoutSpecAttr::get(module.getContext(), kept));
+ }
+
// Emit the llvm.global_ctors array.
buildCtorDtorList(module, cir::CIRDialect::getGlobalCtorsAttrName(),
"llvm.global_ctors", [](mlir::Attribute attr) {
diff --git a/clang/test/CIR/CodeGen/pointer-width-32bit.cpp b/clang/test/CIR/CodeGen/pointer-width-32bit.cpp
new file mode 100644
index 0000000000000..15f3ec92d7e16
--- /dev/null
+++ b/clang/test/CIR/CodeGen/pointer-width-32bit.cpp
@@ -0,0 +1,44 @@
+// RUN: %clang_cc1 -std=c++20 -triple nvptx-nvidia-cuda -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s
+// RUN: %clang_cc1 -std=c++20 -triple nvptx-nvidia-cuda -fclangir -emit-llvm %s -o %t-cir.ll
+// RUN: FileCheck --check-prefix=LLVM --input-file=%t-cir.ll %s
+// RUN: %clang_cc1 -std=c++20 -triple nvptx-nvidia-cuda -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=OGCG --input-file=%t.ll %s
+
+// On a target with 32-bit pointers (e.g. nvptx) both a data pointer (!cir.ptr)
+// and the vtable pointer (!cir.vptr) are 4 bytes wide. The pointer width is
+// carried by a CIR-native data-layout entry keyed on cir.ptr, so the field
+// following a pointer lands at the AST-mandated offset. Sizing pointers as a
+// hardcoded 64 bits previously tripped the record layout builder (insertPadding:
+// assertion `offset >= size`) on every record containing a pointer.
+
+struct S {
+ int *p;
+ int x;
+};
+
+S s;
+
+class A {
+public:
+ virtual void f();
+ int x;
+};
+
+void A::f() {}
+
+// The module carries a CIR-native pointer data-layout entry ({size, abi-align}
+// in bits) that drives both cir.ptr and cir.vptr widths. The 4-byte pointer is
+// immediately followed by 'x' at offset 4 with no padding, and each record is
+// 4-byte aligned.
+// CIR-DAG: !rec_S = !cir.struct<"S" {!cir.ptr<!s32i>, !s32i}>
+// CIR-DAG: !rec_A = !cir.struct<class "A" {!cir.vptr, !s32i}>
+// CIR-DAG: !cir.ptr<!cir.void> = array<i32: 32, 32>
+// CIR: cir.global external @s = #cir.zero : !rec_S {alignment = 4 : i64}
+// CIR: cir.global{{.*}}@_ZTV1A = #cir.vtable<{{.*}}{alignment = 4 : i64}
+
+// LLVM: @s = global %struct.S zeroinitializer, align 4
+// LLVM: @_ZTV1A = global { [3 x ptr] } {{.*}}, align 4
+
+// OGCG: @s = global %struct.S zeroinitializer, align 4
+// OGCG: @_ZTV1A = {{.*}}constant { [3 x ptr] } {{.*}}, align 4
More information about the cfe-commits
mailing list