[clang] Allow enabling sandbox for direct -cc1 invocations

[jansvoboda11]

This PR enables the FS sandbox for direct clang -cc1 invocations. #165350 unintentionally implemented the sandbox only for the code path where clang -cc1 gets invoked after being expanded from a driver command line, which reduced the expected test coverage.

[github-actions]

🪟 Windows x64 Test Results

• 53354 tests passed
• 2217 tests skipped

✅ The build succeeded and all tests passed.

[github-actions]

🐧 Linux x64 Test Results

• 112555 tests passed
• 4598 tests skipped

✅ The build succeeded and all tests passed.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Jan Svoboda (jansvoboda11)

Changes

This PR enables the FS sandbox for direct clang -cc1 invocations. #165350 unintentionally implemented the sandbox only for the code path where clang -cc1 gets invoked after being expanded from a driver command line, which reduced the expected test coverage.

---

Full diff: https://github.com/llvm/llvm-project/pull/174653.diff

2 Files Affected:

• (modified) clang/tools/driver/cc1gen_reproducer_main.cpp (+3)
• (modified) clang/tools/driver/driver.cpp (+8-1)

diff --git a/clang/tools/driver/cc1gen_reproducer_main.cpp b/clang/tools/driver/cc1gen_reproducer_main.cpp
index 14548c39975da..851d252015c44 100644
--- a/clang/tools/driver/cc1gen_reproducer_main.cpp
+++ b/clang/tools/driver/cc1gen_reproducer_main.cpp
@@ -116,6 +116,9 @@ generateReproducerForInvocationArguments(
     ArrayRef<const char *> Argv, const ClangInvocationInfo &Info,
     const llvm::ToolContext &ToolContext,
     IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) {
+  // The driver is not expected to be free of sandbox violations.
+  auto BypassSandbox = llvm::sys::sandbox::scopedDisable();
+
   using namespace driver;
   auto TargetAndMode = ToolChain::getTargetAndModeFromProgramName(Argv[0]);
 
diff --git a/clang/tools/driver/driver.cpp b/clang/tools/driver/driver.cpp
index 1e2c9884ba63d..490136961ebc6 100644
--- a/clang/tools/driver/driver.cpp
+++ b/clang/tools/driver/driver.cpp
@@ -38,6 +38,7 @@
 #include "llvm/Support/CrashRecoveryContext.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/FileSystem.h"
+#include "llvm/Support/IOSandbox.h"
 #include "llvm/Support/LLVMDriver.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/PrettyStackTrace.h"
@@ -264,8 +265,14 @@ int clang_main(int Argc, char **Argv, const llvm::ToolContext &ToolContext) {
   }
 
   // Handle -cc1 integrated tools.
-  if (Args.size() >= 2 && StringRef(Args[1]).starts_with("-cc1"))
+  if (Args.size() >= 2 && StringRef(Args[1]).starts_with("-cc1")) {
+    // Note that this only enables the sandbox for direct -cc1 invocations and
+    // out-of-process -cc1 invocations launched by the driver. For in-process
+    // -cc1 invocations launched by the driver, the sandbox is enabled in
+    // CC1Command::Execute() for better crash recovery.
+    auto EnableSandbox = llvm::sys::sandbox::scopedEnable();
     return ExecuteCC1Tool(Args, ToolContext, VFS);
+  }
 
   // Handle options that need handling before the real command line parsing in
   // Driver::BuildCompilation()

[benlangmuir]

Thanks

[adams381]

Hi @jan-svo,

Thank you for the work on enabling the IO sandbox for direct clang -cc1 invocations. We're working on the ClangIR (CIR) project and discovered that several of our tests are incompatible with LLVM_ENABLE_IO_SANDBOX=ON.

What We're Seeing

15 CIR tests fail with fatal error: error in backend: IO sandbox violation when the sandbox is enabled.

Example test pattern:

clang -cc1 -fclangir -emit-cir -mmlir --mlir-print-ir-before=cir-lowering-prepare \
  test.c -o test.cir 2> before-lp.cir

Error:

fatal error: error in backend: IO sandbox violation
Exit Code: 70

Why We're Using This Pattern

These tests use MLIR's --mlir-print-ir-before=<pass> flag to capture IR at different compilation stages, allowing us to verify that transformations are working correctly. The output goes to stderr, which is redirected to a file for FileCheck verification.

Affected Tests

All 16 CIR tests using --mlir-print-ir-before:

• clang/test/CIR/CodeGen/array-dtor.cpp
• clang/test/CIR/CodeGen/array-ctor.cpp
• clang/test/CIR/CodeGen/dynamic-cast.cpp
• clang/test/CIR/CodeGen/complex-cast.cpp
• clang/test/CIR/CodeGen/complex-mul-div.cpp
• clang/test/CIR/CodeGen/complex-unary.cpp
• clang/test/CIR/CodeGen/binassign.c
• clang/test/CIR/CodeGen/global-ctor-dtor.cpp
• clang/test/CIR/CodeGen/global-init.cpp
• clang/test/CIR/CodeGen/switch_flat_op.cpp
• clang/test/CIR/CodeGen/global-array-dtor.cpp
• clang/test/CIR/CodeGen/pointer-to-member-func.cpp
• clang/test/CIR/CodeGen/pointer-to-data-member-cast.cpp
• clang/test/CIR/CodeGen/global-constant-storage.cpp
• clang/test/CIR/CodeGen/pointer-to-data-member.cpp
• clang/test/CIR/CodeGen/pointer-to-data-member-cmp.cpp

Requesting Guidance

We agree that tests should be sandbox-clean. Could you advise on the best approach to fix this?

Option 1: Should MLIR debug output flags (--mlir-print-ir-before, --mlir-print-ir-after) bypass the sandbox for stderr writes, similar to the bypasses added in #175097 and #175220?

Option 2: Should we restructure these tests to avoid stderr redirection? If so, is there a recommended pattern for verifying IR at intermediate compilation stages?

Option 3: Is there another approach we should consider?

Environment Details

• Default behavior: LLVM_ENABLE_IO_SANDBOX=OFF (tests pass)
• With sandbox enabled: LLVM_ENABLE_IO_SANDBOX=ON (tests fail)
• Commit: 3911c6a2535736a365dd28163dd18daaa4398f74

We want to ensure CIR tests work correctly with the sandbox enabled. Any guidance would be appreciated!