[clang] [AVR] lower `va_arg` in clang rather than rely on the LLVM intrinsic (PR #240157)
Folkert de Vries via cfe-commits
cfe-commits at lists.llvm.org
Sun Nov 8 07:43:57 PST 2026
https://github.com/folkertdev created https://github.com/llvm/llvm-project/pull/240157
The LLVM intrinsic is well-known to miscompile, and most other frontends already lower `va_arg` themselves.
>From cba960bb56ee718b2579c3122182871c69c55ed8 Mon Sep 17 00:00:00 2001
From: Folkert de Vries <folkert at folkertdev.nl>
Date: Sun, 8 Nov 2026 16:32:01 +0100
Subject: [PATCH] [Avr] lower `va_arg` in clang rather than rely on the LLVM
intrinsic
The LLVM intrinsic is well-known to miscompile, and most other frontends already lower va_arg themselves
---
clang/lib/CodeGen/Targets/AVR.cpp | 10 +++
clang/test/CodeGen/avr/avr-vaarg.c | 128 +++++++++++++++++++++++++++++
2 files changed, 138 insertions(+)
create mode 100644 clang/test/CodeGen/avr/avr-vaarg.c
diff --git a/clang/lib/CodeGen/Targets/AVR.cpp b/clang/lib/CodeGen/Targets/AVR.cpp
index 5399d12f7ce805..3d6207d2daa0fc 100644
--- a/clang/lib/CodeGen/Targets/AVR.cpp
+++ b/clang/lib/CodeGen/Targets/AVR.cpp
@@ -107,6 +107,16 @@ class AVRABIInfo : public DefaultABIInfo {
for (auto &I : FI.arguments())
I.info = classifyArgumentType(I.type, NumRegs);
}
+
+ RValue EmitVAArg(CodeGenFunction &CGF, Address VAListAddr, QualType Ty,
+ AggValueSlot Slot) const override {
+ // Variadic arguments are passed directly on the stack, without padding
+ // or alignment.
+ CharUnits SlotSize = CharUnits::One();
+ auto TInfo = getContext().getTypeInfoInChars(Ty);
+ return emitVoidPtrVAArg(CGF, VAListAddr, Ty, /*IsIndirect=*/false, TInfo,
+ SlotSize, /*AllowHigherAlign=*/false, Slot);
+ }
};
class AVRTargetCodeGenInfo : public TargetCodeGenInfo {
diff --git a/clang/test/CodeGen/avr/avr-vaarg.c b/clang/test/CodeGen/avr/avr-vaarg.c
new file mode 100644
index 00000000000000..90f98bffff9854
--- /dev/null
+++ b/clang/test/CodeGen/avr/avr-vaarg.c
@@ -0,0 +1,128 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 6
+// RUN: %clang_cc1 -triple avr -target-cpu atmega328p -emit-llvm -o - %s | FileCheck %s
+#include <stdarg.h>
+
+// CHECK-LABEL: define dso_local i16 @get_int(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 2
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[ARGP_CUR]], align 1
+// CHECK-NEXT: ret i16 [[TMP1]]
+//
+int get_int(va_list *args) {
+ return va_arg(*args, int);
+}
+
+// CHECK-LABEL: define dso_local i32 @get_long(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 4
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARGP_CUR]], align 1
+// CHECK-NEXT: ret i32 [[TMP1]]
+//
+long get_long(va_list *args) {
+ return va_arg(*args, long);
+}
+
+// Values larger than a pointer are still passed directly.
+// CHECK-LABEL: define dso_local i64 @get_long_long(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 8
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr [[ARGP_CUR]], align 1
+// CHECK-NEXT: ret i64 [[TMP1]]
+//
+long long get_long_long(va_list *args) {
+ return va_arg(*args, long long);
+}
+
+// CHECK-LABEL: define dso_local float @get_double(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 4
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[ARGP_CUR]], align 1
+// CHECK-NEXT: ret float [[TMP1]]
+//
+double get_double(va_list *args) {
+ return va_arg(*args, double);
+}
+
+// CHECK-LABEL: define dso_local ptr @get_ptr(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 2
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: [[TMP1:%.*]] = load ptr, ptr [[ARGP_CUR]], align 1
+// CHECK-NEXT: ret ptr [[TMP1]]
+//
+void *get_ptr(va_list *args) {
+ return va_arg(*args, void *);
+}
+
+struct Odd {
+ char a, b, c;
+};
+
+// The pointer is not aligned at all due to SlotSize being 1.
+// CHECK-LABEL: define dso_local %struct.Odd @get_struct_odd(
+// CHECK-SAME: ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[RETVAL:%.*]] = alloca [[STRUCT_ODD:%.*]], align 1
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 3
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: call addrspace(1) void @llvm.memcpy.p0.p0.i16(ptr align 1 [[RETVAL]], ptr align 1 [[ARGP_CUR]], i16 3, i1 false)
+// CHECK-NEXT: [[TMP1:%.*]] = load [[STRUCT_ODD]], ptr [[RETVAL]], align 1
+// CHECK-NEXT: ret [[STRUCT_ODD]] [[TMP1]]
+//
+struct Odd get_struct_odd(va_list *args) {
+ return va_arg(*args, struct Odd);
+}
+
+struct Large {
+ long long a, b;
+};
+
+// Large aggregates are passed directly, not by reference.
+// CHECK-LABEL: define dso_local void @get_struct_large(
+// CHECK-SAME: ptr dead_on_unwind noalias writable sret([[STRUCT_LARGE:%.*]]) align 1 [[AGG_RESULT:%.*]], ptr noundef [[ARGS:%.*]]) addrspace(1) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[ARGS_ADDR:%.*]] = alloca ptr, align 1
+// CHECK-NEXT: store ptr [[ARGS]], ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[ARGS_ADDR]], align 1
+// CHECK-NEXT: [[ARGP_CUR:%.*]] = load ptr, ptr [[TMP0]], align 1
+// CHECK-NEXT: [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i16 16
+// CHECK-NEXT: store ptr [[ARGP_NEXT]], ptr [[TMP0]], align 1
+// CHECK-NEXT: call addrspace(1) void @llvm.memcpy.p0.p0.i16(ptr align 1 [[AGG_RESULT]], ptr align 1 [[ARGP_CUR]], i16 16, i1 false)
+// CHECK-NEXT: ret void
+//
+struct Large get_struct_large(va_list *args) {
+ return va_arg(*args, struct Large);
+}
More information about the cfe-commits
mailing list