[llvm-branch-commits] [clang] [Clang] Enable UBSan for AMDGPU device offload (PR #219973)
Joseph Huber via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Aug 31 18:46:38 PDT 2026
https://github.com/jhuber6 updated https://github.com/llvm/llvm-project/pull/219973
>From ac3631908dc1d1ed29b21df95ed76d7323c8e619 Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Sat, 29 Aug 2026 11:19:57 -0500
Subject: [PATCH 1/4] [Clang] Enable UBSan for AMDGPU device offload
Summary:
This enables the device UBSan runtime for AMDGPU decides. Primarily this
required modifications to the `addSanitizerRuntime` interface so we can
query the compilation's offload status. Also need to forward it through
the linker wrapper interface. Works on all AMDGPU offload, slight hacks
around the other targets as they do not advertise sanitizer
runtimes properly.
This is linked in via a new `-u __ubsan_device_initialize` hook to pull
in the side library. This is standard behavior and keeps the core logic
mostly unchanged and re-used.
---
clang/include/clang/Driver/CommonArgs.h | 2 +-
clang/lib/Driver/ToolChains/AMDGPU.cpp | 2 +-
clang/lib/Driver/ToolChains/Clang.cpp | 3 +-
clang/lib/Driver/ToolChains/CommonArgs.cpp | 35 +++++++++++--
clang/lib/Driver/ToolChains/Cuda.cpp | 2 +-
clang/lib/Driver/ToolChains/Cygwin.cpp | 2 +-
clang/lib/Driver/ToolChains/FreeBSD.cpp | 2 +-
clang/lib/Driver/ToolChains/Fuchsia.cpp | 2 +-
clang/lib/Driver/ToolChains/Gnu.cpp | 2 +-
clang/lib/Driver/ToolChains/Haiku.cpp | 2 +-
clang/lib/Driver/ToolChains/Hexagon.cpp | 4 +-
clang/lib/Driver/ToolChains/NetBSD.cpp | 2 +-
clang/lib/Driver/ToolChains/OpenBSD.cpp | 2 +-
clang/lib/Driver/ToolChains/Solaris.cpp | 2 +-
clang/lib/Driver/ToolChains/UEFI.cpp | 2 +-
.../libclang_rt.ubsan_standalone.a | 0
.../libclang_rt.ubsan_device.a | 0
.../libclang_rt.ubsan_standalone.a | 0
.../fsanitize-undefined-device-offload.c | 50 +++++++++++++++++++
19 files changed, 98 insertions(+), 18 deletions(-)
create mode 100644 clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/amdgpu-amd-amdhsa/libclang_rt.ubsan_standalone.a
create mode 100644 clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_device.a
create mode 100644 clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_standalone.a
create mode 100644 clang/test/Driver/fsanitize-undefined-device-offload.c
diff --git a/clang/include/clang/Driver/CommonArgs.h b/clang/include/clang/Driver/CommonArgs.h
index be15d15a1661e..d7808b978ab6a 100644
--- a/clang/include/clang/Driver/CommonArgs.h
+++ b/clang/include/clang/Driver/CommonArgs.h
@@ -57,7 +57,7 @@ void renderDebugInfoCompressionArgs(const llvm::opt::ArgList &Args,
void claimNoWarnArgs(const llvm::opt::ArgList &Args);
bool addSanitizerRuntimes(const ToolChain &TC, const llvm::opt::ArgList &Args,
- llvm::opt::ArgStringList &CmdArgs);
+ llvm::opt::ArgStringList &CmdArgs, Compilation &C);
void linkSanitizerRuntimeDeps(const ToolChain &TC,
const llvm::opt::ArgList &Args,
diff --git a/clang/lib/Driver/ToolChains/AMDGPU.cpp b/clang/lib/Driver/ToolChains/AMDGPU.cpp
index 6bc6a53732881..d6dacb4ccba83 100644
--- a/clang/lib/Driver/ToolChains/AMDGPU.cpp
+++ b/clang/lib/Driver/ToolChains/AMDGPU.cpp
@@ -667,7 +667,7 @@ void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA,
}
getToolChain().addProfileRTLibs(Args, CmdArgs);
- addSanitizerRuntimes(getToolChain(), Args, CmdArgs);
+ addSanitizerRuntimes(getToolChain(), Args, CmdArgs, C);
if (Args.hasArg(options::OPT_stdlib))
CmdArgs.append({"-lc", "-lm"});
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index 072664e6040f3..c0ee9ab6eff14 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -9889,7 +9889,8 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
return false;
// Don't forward sanitizer arguments if the toolchain doesn't support it.
// Without this check using it on the host would result in linker errors.
- if (requiresUBSanRT(ID) && !ToolChainHasRT(TC, "ubsan_minimal"))
+ if (requiresUBSanRT(ID) && !ToolChainHasRT(TC, "ubsan_minimal") &&
+ !ToolChainHasRT(TC, "ubsan_standalone"))
return false;
// Don't forward -mllvm to toolchains that don't support LLVM.
return TC.HasNativeLLVMSupport() || ID != OPT_mllvm;
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
index 7157c55e9f9e8..1b7904b0242de 100644
--- a/clang/lib/Driver/ToolChains/CommonArgs.cpp
+++ b/clang/lib/Driver/ToolChains/CommonArgs.cpp
@@ -1673,8 +1673,33 @@ void tools::linkSanitizerRuntimeDeps(const ToolChain &TC,
CmdArgs.push_back("-lresolv");
}
+// Host interceptor library for UBSan on the device. Enabled if we are
+// offloading to a target that supports UBSan.
+static bool hostNeedsUbsanDeviceRt(Compilation &C, const ToolChain &HostTC) {
+ if (HostTC.getTriple().isGPU())
+ return false;
+
+ static constexpr Action::OffloadKind Kinds[] = {
+ Action::OFK_Cuda, Action::OFK_OpenMP, Action::OFK_HIP, Action::OFK_SYCL};
+ for (Action::OffloadKind Kind : Kinds) {
+ for (const auto &Entry : llvm::make_range(C.getOffloadToolChains(Kind))) {
+ const ToolChain *DevTC = Entry.second;
+ // FIXME: CUDA/HIPSPV copy the host mask and ignore device sanitizers.
+ if (DevTC->getTriple().isNVPTX() || DevTC->getTriple().isSPIROrSPIRV())
+ continue;
+
+ const ArgList &DevArgs = C.getArgsForToolChain(DevTC, {}, Kind);
+ SanitizerArgs DevSan = DevTC->getSanitizerArgs(DevArgs, {}, Kind);
+ if (DevSan.needsUbsanRt() && !DevSan.requiresMinimalRuntime())
+ return true;
+ }
+ }
+ return false;
+}
+
static void
-collectSanitizerRuntimes(const ToolChain &TC, const ArgList &Args,
+collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
+ const ArgList &Args,
SmallVectorImpl<StringRef> &SharedRuntimes,
SmallVectorImpl<StringRef> &StaticRuntimes,
SmallVectorImpl<StringRef> &NonWholeStaticRuntimes,
@@ -1816,17 +1841,21 @@ collectSanitizerRuntimes(const ToolChain &TC, const ArgList &Args,
}
if (SanArgs.needsUbsanLoopDetectRt())
NonWholeStaticRuntimes.push_back("ubsan_loop_detect");
+ if (hostNeedsUbsanDeviceRt(C, TC)) {
+ NonWholeStaticRuntimes.push_back("ubsan_device");
+ RequiredSymbols.push_back("__ubsan_device_init");
+ }
}
// Should be called before we add system libraries (C++ ABI, libstdc++/libc++,
// C runtime, etc). Returns true if sanitizer system deps need to be linked in.
bool tools::addSanitizerRuntimes(const ToolChain &TC, const ArgList &Args,
- ArgStringList &CmdArgs) {
+ ArgStringList &CmdArgs, Compilation &C) {
const SanitizerArgs &SanArgs = TC.getSanitizerArgs(Args);
SmallVector<StringRef, 4> SharedRuntimes, StaticRuntimes,
NonWholeStaticRuntimes, HelperStaticRuntimes, RequiredSymbols;
if (SanArgs.linkRuntimes()) {
- collectSanitizerRuntimes(TC, Args, SharedRuntimes, StaticRuntimes,
+ collectSanitizerRuntimes(C, TC, Args, SharedRuntimes, StaticRuntimes,
NonWholeStaticRuntimes, HelperStaticRuntimes,
RequiredSymbols);
}
diff --git a/clang/lib/Driver/ToolChains/Cuda.cpp b/clang/lib/Driver/ToolChains/Cuda.cpp
index 7ad712adf3eb4..d8fdbff81a992 100644
--- a/clang/lib/Driver/ToolChains/Cuda.cpp
+++ b/clang/lib/Driver/ToolChains/Cuda.cpp
@@ -672,7 +672,7 @@ void NVPTX::Linker::ConstructJob(Compilation &C, const JobAction &JA,
CmdArgs.push_back(Args.MakeArgString(Twine("-L") + DefaultLibPath));
getToolChain().addProfileRTLibs(Args, CmdArgs);
- addSanitizerRuntimes(getToolChain(), Args, CmdArgs);
+ addSanitizerRuntimes(getToolChain(), Args, CmdArgs, C);
if (Args.hasArg(options::OPT_stdlib))
CmdArgs.append({"-lc", "-lm"});
diff --git a/clang/lib/Driver/ToolChains/Cygwin.cpp b/clang/lib/Driver/ToolChains/Cygwin.cpp
index 1a9f32b4586ce..95ce92ef9eb8c 100644
--- a/clang/lib/Driver/ToolChains/Cygwin.cpp
+++ b/clang/lib/Driver/ToolChains/Cygwin.cpp
@@ -277,7 +277,7 @@ void cygwin::Linker::ConstructJob(Compilation &C, const JobAction &JA,
CmdArgs.push_back("--disable-nxcompat");
bool NeedsSanitizerDeps =
- tools::addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ tools::addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
bool NeedsXRayDeps = tools::addXRayRuntime(ToolChain, Args, CmdArgs);
tools::addLinkerCompressDebugSectionsOption(ToolChain, Args, CmdArgs);
tools::AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/FreeBSD.cpp b/clang/lib/Driver/ToolChains/FreeBSD.cpp
index dcf2973c364d1..273cb3bb89607 100644
--- a/clang/lib/Driver/ToolChains/FreeBSD.cpp
+++ b/clang/lib/Driver/ToolChains/FreeBSD.cpp
@@ -275,7 +275,7 @@ void freebsd::Linker::ConstructJob(Compilation &C, const JobAction &JA,
if (auto LTO = ToolChain.getLTOMode(Args); LTO != LTOK_None)
addLTOOptions(ToolChain, Args, CmdArgs, Output, Inputs, LTO == LTOK_Thin);
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
bool NeedsXRayDeps = addXRayRuntime(ToolChain, Args, CmdArgs);
addLinkerCompressDebugSectionsOption(ToolChain, Args, CmdArgs);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/Fuchsia.cpp b/clang/lib/Driver/ToolChains/Fuchsia.cpp
index 028d94b3fa103..3b5729b45e7e8 100644
--- a/clang/lib/Driver/ToolChains/Fuchsia.cpp
+++ b/clang/lib/Driver/ToolChains/Fuchsia.cpp
@@ -173,7 +173,7 @@ void fuchsia::Linker::ConstructJob(Compilation &C, const JobAction &JA,
// Note that Fuchsia never needs to link in sanitizer runtime deps. Any
// sanitizer runtimes with system dependencies use the `.deplibs` feature
// instead.
- addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
addXRayRuntime(ToolChain, Args, CmdArgs);
diff --git a/clang/lib/Driver/ToolChains/Gnu.cpp b/clang/lib/Driver/ToolChains/Gnu.cpp
index 04f3b8d2200d6..d2c98aad865fe 100644
--- a/clang/lib/Driver/ToolChains/Gnu.cpp
+++ b/clang/lib/Driver/ToolChains/Gnu.cpp
@@ -445,7 +445,7 @@ void tools::gnutools::Linker::ConstructJob(Compilation &C, const JobAction &JA,
if (Args.hasArg(options::OPT_Z_Xlinker__no_demangle))
CmdArgs.push_back("--no-demangle");
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
bool NeedsXRayDeps = addXRayRuntime(ToolChain, Args, CmdArgs);
addLinkerCompressDebugSectionsOption(ToolChain, Args, CmdArgs);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/Haiku.cpp b/clang/lib/Driver/ToolChains/Haiku.cpp
index a155f22b75a79..36b07f888cc66 100644
--- a/clang/lib/Driver/ToolChains/Haiku.cpp
+++ b/clang/lib/Driver/ToolChains/Haiku.cpp
@@ -90,7 +90,7 @@ void haiku::Linker::ConstructJob(Compilation &C, const JobAction &JA,
if (auto LTO = ToolChain.getLTOMode(Args); LTO != LTOK_None)
addLTOOptions(ToolChain, Args, CmdArgs, Output, Inputs, LTO == LTOK_Thin);
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
addLinkerCompressDebugSectionsOption(ToolChain, Args, CmdArgs);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/Hexagon.cpp b/clang/lib/Driver/ToolChains/Hexagon.cpp
index 76c5cafd77bc9..933b9f2ea5412 100644
--- a/clang/lib/Driver/ToolChains/Hexagon.cpp
+++ b/clang/lib/Driver/ToolChains/Hexagon.cpp
@@ -218,7 +218,7 @@ void hexagon::Assembler::ConstructJob(Compilation &C, const JobAction &JA,
"-mcpu=hexagon" +
toolchains::HexagonToolChain::GetTargetCPUVersion(Args)));
- addSanitizerRuntimes(HTC, Args, CmdArgs);
+ addSanitizerRuntimes(HTC, Args, CmdArgs, C);
assert((Output.isFilename() || Output.isNothing()) && "Invalid output.");
if (Output.isFilename()) {
@@ -304,7 +304,7 @@ constructHexagonLinkArgs(Compilation &C, const JobAction &JA,
bool UseShared = IsShared && !IsStatic;
StringRef CpuVer = toolchains::HexagonToolChain::GetTargetCPUVersion(Args);
- bool NeedsSanitizerDeps = addSanitizerRuntimes(HTC, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(HTC, Args, CmdArgs, C);
bool NeedsXRayDeps = addXRayRuntime(HTC, Args, CmdArgs);
//----------------------------------------------------------------------------
diff --git a/clang/lib/Driver/ToolChains/NetBSD.cpp b/clang/lib/Driver/ToolChains/NetBSD.cpp
index f03114b53bb61..dd53606351b24 100644
--- a/clang/lib/Driver/ToolChains/NetBSD.cpp
+++ b/clang/lib/Driver/ToolChains/NetBSD.cpp
@@ -274,7 +274,7 @@ void netbsd::Linker::ConstructJob(Compilation &C, const JobAction &JA,
options::OPT_s, options::OPT_t});
ToolChain.AddFilePathLibArgs(Args, CmdArgs);
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
bool NeedsXRayDeps = addXRayRuntime(ToolChain, Args, CmdArgs);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/OpenBSD.cpp b/clang/lib/Driver/ToolChains/OpenBSD.cpp
index fa36726534bed..d1c00898bd425 100644
--- a/clang/lib/Driver/ToolChains/OpenBSD.cpp
+++ b/clang/lib/Driver/ToolChains/OpenBSD.cpp
@@ -206,7 +206,7 @@ void openbsd::Linker::ConstructJob(Compilation &C, const JobAction &JA,
if (auto LTO = ToolChain.getLTOMode(Args); LTO != LTOK_None)
addLTOOptions(ToolChain, Args, CmdArgs, Output, Inputs, LTO == LTOK_Thin);
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
bool NeedsXRayDeps = addXRayRuntime(ToolChain, Args, CmdArgs);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
diff --git a/clang/lib/Driver/ToolChains/Solaris.cpp b/clang/lib/Driver/ToolChains/Solaris.cpp
index e5f1cbb033c3b..8eafe18a1eb30 100644
--- a/clang/lib/Driver/ToolChains/Solaris.cpp
+++ b/clang/lib/Driver/ToolChains/Solaris.cpp
@@ -202,7 +202,7 @@ void solaris::Linker::ConstructJob(Compilation &C, const JobAction &JA,
Args.addAllArgs(CmdArgs, {options::OPT_L, options::OPT_T_Group});
- bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs);
+ bool NeedsSanitizerDeps = addSanitizerRuntimes(ToolChain, Args, CmdArgs, C);
AddLinkerInputs(ToolChain, Inputs, Args, CmdArgs, JA);
if (!Args.hasArg(options::OPT_nostdlib, options::OPT_nodefaultlibs,
diff --git a/clang/lib/Driver/ToolChains/UEFI.cpp b/clang/lib/Driver/ToolChains/UEFI.cpp
index e26403ebe9030..6e93f7c4be193 100644
--- a/clang/lib/Driver/ToolChains/UEFI.cpp
+++ b/clang/lib/Driver/ToolChains/UEFI.cpp
@@ -90,7 +90,7 @@ void tools::uefi::Linker::ConstructJob(Compilation &C, const JobAction &JA,
bool NoLibc = Args.hasArg(options::OPT_nolibc);
if (!Args.hasArg(options::OPT_nostdlib, options::OPT_nodefaultlibs,
options::OPT_r)) {
- addSanitizerRuntimes(TC, Args, CmdArgs);
+ addSanitizerRuntimes(TC, Args, CmdArgs, C);
addXRayRuntime(TC, Args, CmdArgs);
diff --git a/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/amdgpu-amd-amdhsa/libclang_rt.ubsan_standalone.a b/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/amdgpu-amd-amdhsa/libclang_rt.ubsan_standalone.a
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_device.a b/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_device.a
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_standalone.a b/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_standalone.a
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/clang/test/Driver/fsanitize-undefined-device-offload.c b/clang/test/Driver/fsanitize-undefined-device-offload.c
new file mode 100644
index 0000000000000..36772e4606b7d
--- /dev/null
+++ b/clang/test/Driver/fsanitize-undefined-device-offload.c
@@ -0,0 +1,50 @@
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -fopenmp=libomp --offload-arch=gfx908 -fsanitize=undefined -nogpuinc \
+// RUN: --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-OPENMP
+// CHECK-OPENMP-DAG: "--device-compiler=amdgpu-amd-amdhsa=-fsanitize=undefined"
+// CHECK-OPENMP-DAG: "-u" "__ubsan_device_init"
+// CHECK-OPENMP-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -fsanitize=undefined -nogpuinc -nogpulib \
+// RUN: --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-HIP-HOST
+// CHECK-HIP-HOST-DAG: "-u" "__ubsan_device_init"
+// CHECK-HIP-HOST-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -Xarch_device -fsanitize=undefined \
+// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-DEV
+// CHECK-XARCH-DEV-DAG: "-u" "__ubsan_device_init"
+// CHECK-XARCH-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -Xarch_host -fsanitize=undefined \
+// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-HOST
+// CHECK-XARCH-HOST-NOT: ubsan_device
+// CHECK-XARCH-HOST-NOT: __ubsan_device_init
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -fsanitize=undefined \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-HOST
+// CHECK-HOST-NOT: ubsan_device
+// CHECK-HOST-NOT: __ubsan_device_init
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -fsanitize=undefined \
+// RUN: -fsanitize-minimal-runtime -nogpuinc -nogpulib \
+// RUN: --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-MINIMAL
+// CHECK-MINIMAL-NOT: ubsan_device
+// CHECK-MINIMAL-NOT: __ubsan_device_init
+
+int main(void) { return 0; }
>From 18f42126424af9efd8329d1bb185e73f34e9341f Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Mon, 31 Aug 2026 16:45:45 -0500
Subject: [PATCH 2/4] [Clang] Link libclang_rt.ubsan_offload.a for device
offload
Match the compiler-rt rename of the host interceptor and the
-u __ubsan_offload_init hook.
---
clang/lib/Driver/ToolChains/CommonArgs.cpp | 10 ++++----
...n_device.a => libclang_rt.ubsan_offload.a} | 0
...ffload.c => fsanitize-undefined-offload.c} | 24 +++++++++----------
3 files changed, 17 insertions(+), 17 deletions(-)
rename clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/{libclang_rt.ubsan_device.a => libclang_rt.ubsan_offload.a} (100%)
rename clang/test/Driver/{fsanitize-undefined-device-offload.c => fsanitize-undefined-offload.c} (83%)
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
index 1b7904b0242de..60ff873e58de8 100644
--- a/clang/lib/Driver/ToolChains/CommonArgs.cpp
+++ b/clang/lib/Driver/ToolChains/CommonArgs.cpp
@@ -1673,9 +1673,9 @@ void tools::linkSanitizerRuntimeDeps(const ToolChain &TC,
CmdArgs.push_back("-lresolv");
}
-// Host interceptor library for UBSan on the device. Enabled if we are
+// Host interceptor library for offload UBSan. Enabled if we are
// offloading to a target that supports UBSan.
-static bool hostNeedsUbsanDeviceRt(Compilation &C, const ToolChain &HostTC) {
+static bool hostNeedsUbsanOffloadRt(Compilation &C, const ToolChain &HostTC) {
if (HostTC.getTriple().isGPU())
return false;
@@ -1841,9 +1841,9 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
}
if (SanArgs.needsUbsanLoopDetectRt())
NonWholeStaticRuntimes.push_back("ubsan_loop_detect");
- if (hostNeedsUbsanDeviceRt(C, TC)) {
- NonWholeStaticRuntimes.push_back("ubsan_device");
- RequiredSymbols.push_back("__ubsan_device_init");
+ if (hostNeedsUbsanOffloadRt(C, TC)) {
+ NonWholeStaticRuntimes.push_back("ubsan_offload");
+ RequiredSymbols.push_back("__ubsan_offload_init");
}
}
diff --git a/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_device.a b/clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_offload.a
similarity index 100%
rename from clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_device.a
rename to clang/test/Driver/Inputs/resource_dir_with_amdgpu_per_target_subdir/lib/x86_64-unknown-linux-gnu/libclang_rt.ubsan_offload.a
diff --git a/clang/test/Driver/fsanitize-undefined-device-offload.c b/clang/test/Driver/fsanitize-undefined-offload.c
similarity index 83%
rename from clang/test/Driver/fsanitize-undefined-device-offload.c
rename to clang/test/Driver/fsanitize-undefined-offload.c
index 36772e4606b7d..49f8b8d506bf2 100644
--- a/clang/test/Driver/fsanitize-undefined-device-offload.c
+++ b/clang/test/Driver/fsanitize-undefined-offload.c
@@ -4,39 +4,39 @@
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-OPENMP
// CHECK-OPENMP-DAG: "--device-compiler=amdgpu-amd-amdhsa=-fsanitize=undefined"
-// CHECK-OPENMP-DAG: "-u" "__ubsan_device_init"
-// CHECK-OPENMP-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+// CHECK-OPENMP-DAG: "-u" "__ubsan_offload_init"
+// CHECK-OPENMP-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -x hip --offload-arch=gfx908 -fsanitize=undefined -nogpuinc -nogpulib \
// RUN: --rocm-path=%S/Inputs/rocm \
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-HIP-HOST
-// CHECK-HIP-HOST-DAG: "-u" "__ubsan_device_init"
-// CHECK-HIP-HOST-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+// CHECK-HIP-HOST-DAG: "-u" "__ubsan_offload_init"
+// CHECK-HIP-HOST-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -x hip --offload-arch=gfx908 -Xarch_device -fsanitize=undefined \
// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-DEV
-// CHECK-XARCH-DEV-DAG: "-u" "__ubsan_device_init"
-// CHECK-XARCH-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_device.a"
+// CHECK-XARCH-DEV-DAG: "-u" "__ubsan_offload_init"
+// CHECK-XARCH-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -x hip --offload-arch=gfx908 -Xarch_host -fsanitize=undefined \
// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-HOST
-// CHECK-XARCH-HOST-NOT: ubsan_device
-// CHECK-XARCH-HOST-NOT: __ubsan_device_init
+// CHECK-XARCH-HOST-NOT: ubsan_offload
+// CHECK-XARCH-HOST-NOT: __ubsan_offload_init
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -fsanitize=undefined \
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-HOST
-// CHECK-HOST-NOT: ubsan_device
-// CHECK-HOST-NOT: __ubsan_device_init
+// CHECK-HOST-NOT: ubsan_offload
+// CHECK-HOST-NOT: __ubsan_offload_init
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -x hip --offload-arch=gfx908 -fsanitize=undefined \
@@ -44,7 +44,7 @@
// RUN: --rocm-path=%S/Inputs/rocm \
// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
// RUN: | FileCheck %s --check-prefix=CHECK-MINIMAL
-// CHECK-MINIMAL-NOT: ubsan_device
-// CHECK-MINIMAL-NOT: __ubsan_device_init
+// CHECK-MINIMAL-NOT: ubsan_offload
+// CHECK-MINIMAL-NOT: __ubsan_offload_init
int main(void) { return 0; }
>From e78d18f75ba5167c99c4bf25c92017f2e7e810f3 Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Mon, 31 Aug 2026 17:45:41 -0500
Subject: [PATCH 3/4] Device only and -shared fix
---
clang/lib/Driver/ToolChains/CommonArgs.cpp | 24 ++++++++++-------
.../test/Driver/fsanitize-undefined-offload.c | 26 +++++++++++++++++++
2 files changed, 41 insertions(+), 9 deletions(-)
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
index 60ff873e58de8..2aff00dd0404a 100644
--- a/clang/lib/Driver/ToolChains/CommonArgs.cpp
+++ b/clang/lib/Driver/ToolChains/CommonArgs.cpp
@@ -1673,8 +1673,7 @@ void tools::linkSanitizerRuntimeDeps(const ToolChain &TC,
CmdArgs.push_back("-lresolv");
}
-// Host interceptor library for offload UBSan. Enabled if we are
-// offloading to a target that supports UBSan.
+// Host interceptor library for offload UBSan.
static bool hostNeedsUbsanOffloadRt(Compilation &C, const ToolChain &HostTC) {
if (HostTC.getTriple().isGPU())
return false;
@@ -1685,7 +1684,8 @@ static bool hostNeedsUbsanOffloadRt(Compilation &C, const ToolChain &HostTC) {
for (const auto &Entry : llvm::make_range(C.getOffloadToolChains(Kind))) {
const ToolChain *DevTC = Entry.second;
// FIXME: CUDA/HIPSPV copy the host mask and ignore device sanitizers.
- if (DevTC->getTriple().isNVPTX() || DevTC->getTriple().isSPIROrSPIRV())
+ const llvm::Triple &TT = DevTC->getTriple();
+ if (!TT.isAMDGCN() || TT.getOS() != llvm::Triple::AMDHSA)
continue;
const ArgList &DevArgs = C.getArgsForToolChain(DevTC, {}, Kind);
@@ -1707,6 +1707,8 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
SmallVectorImpl<StringRef> &RequiredSymbols) {
assert(!TC.getTriple().isOSDarwin() && "it's not used by Darwin");
const SanitizerArgs &SanArgs = TC.getSanitizerArgs(Args);
+ const bool NeedsOffloadRt = hostNeedsUbsanOffloadRt(C, TC);
+ const bool NeedsUbsanRt = SanArgs.needsUbsanRt() || NeedsOffloadRt;
// Collect shared runtimes.
if (SanArgs.needsSharedRt()) {
if (SanArgs.needsAsanRt()) {
@@ -1721,7 +1723,7 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
}
if (SanArgs.needsNsanRt())
SharedRuntimes.push_back("nsan");
- if (SanArgs.needsUbsanRt()) {
+ if (NeedsUbsanRt) {
if (SanArgs.requiresMinimalRuntime())
SharedRuntimes.push_back("ubsan_minimal");
else
@@ -1754,9 +1756,17 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
if (SanArgs.needsAsanRt())
HelperStaticRuntimes.push_back("asan_static");
+ // Offloading images can live in DSOs, the host interceptors must follow.
+ if (NeedsOffloadRt) {
+ NonWholeStaticRuntimes.push_back("ubsan_offload");
+ RequiredSymbols.push_back("__ubsan_offload_init");
+ }
+
// Collect static runtimes.
if (Args.hasArg(options::OPT_shared)) {
// Don't link static runtimes into DSOs.
+ if (NeedsOffloadRt && !SanArgs.needsSharedRt() && !SanArgs.needsUbsanRt())
+ StaticRuntimes.push_back("ubsan_standalone");
return;
}
@@ -1808,7 +1818,7 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
}
if (!SanArgs.needsSharedRt() && SanArgs.needsTysanRt())
StaticRuntimes.push_back("tysan");
- if (!SanArgs.needsSharedRt() && SanArgs.needsUbsanRt()) {
+ if (!SanArgs.needsSharedRt() && NeedsUbsanRt) {
if (SanArgs.requiresMinimalRuntime()) {
StaticRuntimes.push_back("ubsan_minimal");
} else {
@@ -1841,10 +1851,6 @@ collectSanitizerRuntimes(Compilation &C, const ToolChain &TC,
}
if (SanArgs.needsUbsanLoopDetectRt())
NonWholeStaticRuntimes.push_back("ubsan_loop_detect");
- if (hostNeedsUbsanOffloadRt(C, TC)) {
- NonWholeStaticRuntimes.push_back("ubsan_offload");
- RequiredSymbols.push_back("__ubsan_offload_init");
- }
}
// Should be called before we add system libraries (C++ ABI, libstdc++/libc++,
diff --git a/clang/test/Driver/fsanitize-undefined-offload.c b/clang/test/Driver/fsanitize-undefined-offload.c
index 49f8b8d506bf2..bed6cbbd58972 100644
--- a/clang/test/Driver/fsanitize-undefined-offload.c
+++ b/clang/test/Driver/fsanitize-undefined-offload.c
@@ -22,6 +22,7 @@
// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-DEV
// CHECK-XARCH-DEV-DAG: "-u" "__ubsan_offload_init"
// CHECK-XARCH-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
+// CHECK-XARCH-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_standalone.a"
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -x hip --offload-arch=gfx908 -Xarch_host -fsanitize=undefined \
@@ -47,4 +48,29 @@
// CHECK-MINIMAL-NOT: ubsan_offload
// CHECK-MINIMAL-NOT: __ubsan_offload_init
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -Xarch_device -fsanitize=undefined \
+// RUN: -fPIC -shared -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-SHARED-DEV
+// CHECK-SHARED-DEV-DAG: "-u" "__ubsan_offload_init"
+// CHECK-SHARED-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
+// CHECK-SHARED-DEV-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_standalone.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -fsanitize=undefined \
+// RUN: -fPIC -shared -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-SHARED
+// CHECK-SHARED-DAG: "-u" "__ubsan_offload_init"
+// CHECK-SHARED-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -fopenmp=libomp -fopenmp-targets=x86_64-unknown-linux-gnu \
+// RUN: -fsanitize=undefined \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-OMP-CPU
+// CHECK-OMP-CPU-NOT: ubsan_offload
+// CHECK-OMP-CPU-NOT: __ubsan_offload_init
+
int main(void) { return 0; }
>From 01cc74d094ac1fec081313c0a5bc613e16a87140 Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Mon, 31 Aug 2026 20:46:03 -0500
Subject: [PATCH 4/4] [Clang] Honor -Xarch_gfx* when linking the UBSan offload
runtime
Empty bound architecture misses per-GPU sanitizer flags, so inspect each
offload arch when deciding whether the host interceptor is required.
---
clang/lib/Driver/ToolChains/CommonArgs.cpp | 11 +++++++----
clang/test/Driver/fsanitize-undefined-offload.c | 16 ++++++++++++++++
2 files changed, 23 insertions(+), 4 deletions(-)
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
index 2aff00dd0404a..3af6593ae2360 100644
--- a/clang/lib/Driver/ToolChains/CommonArgs.cpp
+++ b/clang/lib/Driver/ToolChains/CommonArgs.cpp
@@ -1688,10 +1688,13 @@ static bool hostNeedsUbsanOffloadRt(Compilation &C, const ToolChain &HostTC) {
if (!TT.isAMDGCN() || TT.getOS() != llvm::Triple::AMDHSA)
continue;
- const ArgList &DevArgs = C.getArgsForToolChain(DevTC, {}, Kind);
- SanitizerArgs DevSan = DevTC->getSanitizerArgs(DevArgs, {}, Kind);
- if (DevSan.needsUbsanRt() && !DevSan.requiresMinimalRuntime())
- return true;
+ for (BoundArch BA :
+ C.getDriver().getOffloadArchs(C, C.getArgs(), Kind, *DevTC)) {
+ const ArgList &DevArgs = C.getArgsForToolChain(DevTC, BA, Kind);
+ SanitizerArgs DevSan = DevTC->getSanitizerArgs(DevArgs, BA, Kind);
+ if (DevSan.needsUbsanRt() && !DevSan.requiresMinimalRuntime())
+ return true;
+ }
}
}
return false;
diff --git a/clang/test/Driver/fsanitize-undefined-offload.c b/clang/test/Driver/fsanitize-undefined-offload.c
index bed6cbbd58972..d4fd4462d7db7 100644
--- a/clang/test/Driver/fsanitize-undefined-offload.c
+++ b/clang/test/Driver/fsanitize-undefined-offload.c
@@ -65,6 +65,22 @@
// CHECK-SHARED-DAG: "-u" "__ubsan_offload_init"
// CHECK-SHARED-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -Xarch_gfx908 -fsanitize=undefined \
+// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-GPU
+// CHECK-XARCH-GPU-DAG: "-u" "__ubsan_offload_init"
+// CHECK-XARCH-GPU-DAG: "{{[^"]*}}x86_64-unknown-linux-gnu{{/|\\\\}}libclang_rt.ubsan_offload.a"
+
+// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
+// RUN: -x hip --offload-arch=gfx908 -Xarch_gfx90a -fsanitize=undefined \
+// RUN: -nogpuinc -nogpulib --rocm-path=%S/Inputs/rocm \
+// RUN: -resource-dir=%S/Inputs/resource_dir_with_amdgpu_per_target_subdir %s 2>&1 \
+// RUN: | FileCheck %s --check-prefix=CHECK-XARCH-OTHER
+// CHECK-XARCH-OTHER-NOT: ubsan_offload
+// CHECK-XARCH-OTHER-NOT: __ubsan_offload_init
+
// RUN: %clang -no-canonical-prefixes -### --target=x86_64-unknown-linux-gnu \
// RUN: -fopenmp=libomp -fopenmp-targets=x86_64-unknown-linux-gnu \
// RUN: -fsanitize=undefined \
More information about the llvm-branch-commits
mailing list