[Polly] Forward VFS from PassBuilder for IO sandboxing (#188657)

#184545 default-enables the IO sandbox in assert-builds. This causes
Clang using Polly to crash (#188568).

The issue is that `PassBuilder` uses `vfs::getRealFileSystem()` by
default which is considered a IO sandbox violation in the Clang process.
With this PR store the VFS from the `PassBuilder` from the original
`registerPollyPasses` call for creating other `PassBuilder` instances.

This PR also adds infrastructure for running Polly in `clang` (in
addition in `opt`). `opt` does not enable the sandbox such that we need
separate tests using Clang.

Closes: #188568

polly/ci/polly-x86_64-linux-plugin.py

@@ -29,7 +29,7 @@ with worker.run(
             "-DCMAKE_BUILD_TYPE=Release",
             "-DCMAKE_C_COMPILER_LAUNCHER=ccache",
             "-DCMAKE_CXX_COMPILER_LAUNCHER=ccache",
-            "-DLLVM_ENABLE_PROJECTS=polly",
+            "-DLLVM_ENABLE_PROJECTS=polly;clang",
             "-DLLVM_TARGETS_TO_BUILD=X86",
             "-DLLVM_ENABLE_LLD=ON",
             "-DLLVM_ENABLE_ASSERTIONS=ON",

polly/ci/polly-x86_64-linux.py

@@ -30,7 +30,7 @@ with worker.run(
             "-DCMAKE_BUILD_TYPE=Release",
             "-DCMAKE_C_COMPILER_LAUNCHER=ccache",
             "-DCMAKE_CXX_COMPILER_LAUNCHER=ccache",
-            "-DLLVM_ENABLE_PROJECTS=polly",
+            "-DLLVM_ENABLE_PROJECTS=polly;clang",
             "-DLLVM_TARGETS_TO_BUILD=X86",
             "-DLLVM_ENABLE_LLD=ON",
             "-DLLVM_ENABLE_ASSERTIONS=ON",

polly/include/polly/ScopInliner.h

@@ -9,14 +9,23 @@
 #ifndef POLLY_POLLYINLINER_H
 #define POLLY_POLLYINLINER_H
 
+#include "llvm/ADT/IntrusiveRefCntPtr.h"
 #include "llvm/Analysis/CGSCCPassManager.h"
 #include "llvm/Analysis/LazyCallGraph.h"
 #include "llvm/IR/PassManager.h"
 
+namespace llvm {
+namespace vfs {
+class FileSystem;
+}
+} // namespace llvm
+
 namespace polly {
 class ScopInlinerPass : public llvm::PassInfoMixin<ScopInlinerPass> {
+  llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> FS;
+
 public:
-  ScopInlinerPass();
+  explicit ScopInlinerPass(llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> FS);
 
   llvm::PreservedAnalyses run(llvm::LazyCallGraph::SCC &C,
                               llvm::CGSCCAnalysisManager &AM,

polly/lib/Support/PollyPasses.def

@@ -8,7 +8,7 @@ MODULE_PASS("polly-custom", createModuleToFunctionPassAdaptor(PollyFunctionPass(
 #ifndef CGSCC_PASS
 #define CGSCC_PASS(NAME, CREATE_PASS, PARSER)
 #endif
-CGSCC_PASS("polly-inline", ScopInlinerPass(), parseNoOptions)
+CGSCC_PASS("polly-inline", ScopInlinerPass(FS), parseNoOptions)
 #undef CGSCC_PASS
 
 #ifndef FUNCTION_PASS

polly/lib/Support/RegisterPasses.cpp

@@ -459,8 +459,13 @@ parsePollyCustomOptions(StringRef Params) {
 ///                     The IR may still be modified.
 static void buildCommonPollyPipeline(FunctionPassManager &PM,
                                      OptimizationLevel Level,
+                                     IntrusiveRefCntPtr<vfs::FileSystem> FS,
                                      bool EnableForOpt) {
-  PassBuilder PB;
+  PassBuilder PB(
+      /*TM=*/nullptr,
+      /*PipelineTuningOptions=*/{},
+      /*PGOOpt=*/{},
+      /*PIC=*/nullptr, std::move(FS));
 
   ExitOnError Err("Inconsistent Polly configuration: ");
   PollyPassOptions &&Opts =
@@ -475,7 +480,8 @@ static void buildCommonPollyPipeline(FunctionPassManager &PM,
 }
 
 static void buildEarlyPollyPipeline(llvm::ModulePassManager &MPM,
-                                    llvm::OptimizationLevel Level) {
+                                    llvm::OptimizationLevel Level,
+                                    IntrusiveRefCntPtr<vfs::FileSystem> FS) {
   bool EnableForOpt =
       shouldEnablePollyForOptimization() && Level.isOptimizingForSpeed();
   if (!shouldEnablePollyForDiagnostic() && !EnableForOpt)
@@ -494,7 +500,7 @@ static void buildEarlyPollyPipeline(llvm::ModulePassManager &MPM,
     FPM = FunctionPassManager();
   }
 
-  buildCommonPollyPipeline(FPM, Level, EnableForOpt);
+  buildCommonPollyPipeline(FPM, Level, std::move(FS), EnableForOpt);
   MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
 
   if (DumpAfter)
@@ -504,7 +510,8 @@ static void buildEarlyPollyPipeline(llvm::ModulePassManager &MPM,
 }
 
 static void buildLatePollyPipeline(FunctionPassManager &PM,
-                                   llvm::OptimizationLevel Level) {
+                                   llvm::OptimizationLevel Level,
+                                   IntrusiveRefCntPtr<vfs::FileSystem> FS) {
   bool EnableForOpt =
       shouldEnablePollyForOptimization() && Level.isOptimizingForSpeed();
   if (!shouldEnablePollyForDiagnostic() && !EnableForOpt)
@@ -518,7 +525,7 @@ static void buildLatePollyPipeline(FunctionPassManager &PM,
         "not supported with NPM",
         false);
 
-  buildCommonPollyPipeline(PM, Level, EnableForOpt);
+  buildCommonPollyPipeline(PM, Level, std::move(FS), EnableForOpt);
 
   if (DumpAfter)
     PM.addPass(DumpFunctionPass("-after"));
@@ -542,7 +549,8 @@ static llvm::Expected<std::monostate> parseNoOptions(StringRef Params) {
 static llvm::Expected<bool>
 parseCGPipeline(StringRef Name, llvm::CGSCCPassManager &CGPM,
                 PassInstrumentationCallbacks *PIC,
-                ArrayRef<PassBuilder::PipelineElement> Pipeline) {
+                ArrayRef<PassBuilder::PipelineElement> Pipeline,
+                IntrusiveRefCntPtr<vfs::FileSystem> FS) {
 #define CGSCC_PASS(NAME, CREATE_PASS, PARSER)                                  \
   if (PassBuilder::checkParametrizedPassName(Name, NAME)) {                    \
     auto Params = PassBuilder::parsePassParameters(PARSER, Name, NAME);        \
@@ -624,6 +632,7 @@ parseModulePipeline(StringRef Name, llvm::ModulePassManager &MPM,
 /// handle LICMed code to make it useful.
 void registerPollyPasses(PassBuilder &PB) {
   PassInstrumentationCallbacks *PIC = PB.getPassInstrumentationCallbacks();
+  IntrusiveRefCntPtr<vfs::FileSystem> FS = PB.getVirtualFileSystemPtr();
 
 #define MODULE_PASS(NAME, CREATE_PASS, PARSER)                                 \
   {                                                                            \
@@ -652,10 +661,10 @@ void registerPollyPasses(PassBuilder &PB) {
         return Err(parseFunctionPipeline(Name, FPM, PIC, Pipeline));
       });
   PB.registerPipelineParsingCallback(
-      [PIC](StringRef Name, CGSCCPassManager &CGPM,
-            ArrayRef<PassBuilder::PipelineElement> Pipeline) -> bool {
+      [PIC, FS](StringRef Name, CGSCCPassManager &CGPM,
+                ArrayRef<PassBuilder::PipelineElement> Pipeline) -> bool {
         ExitOnError Err("Unable to parse Polly call graph pass: ");
-        return Err(parseCGPipeline(Name, CGPM, PIC, Pipeline));
+        return Err(parseCGPipeline(Name, CGPM, PIC, Pipeline, FS));
       });
   PB.registerPipelineParsingCallback(
       [PIC](StringRef Name, ModulePassManager &MPM,
@@ -666,10 +675,16 @@ void registerPollyPasses(PassBuilder &PB) {
 
   switch (PassPosition) {
   case POSITION_EARLY:
-    PB.registerPipelineStartEPCallback(buildEarlyPollyPipeline);
+    PB.registerPipelineStartEPCallback(
+        [FS](ModulePassManager &MPM, OptimizationLevel Level) {
+          buildEarlyPollyPipeline(MPM, Level, FS);
+        });
     break;
   case POSITION_BEFORE_VECTORIZER:
-    PB.registerVectorizerStartEPCallback(buildLatePollyPipeline);
+    PB.registerVectorizerStartEPCallback(
+        [FS](FunctionPassManager &FPM, OptimizationLevel Level) {
+          buildLatePollyPipeline(FPM, Level, FS);
+        });
     break;
   }
 }

polly/lib/Transform/ScopInliner.cpp

@@ -34,7 +34,9 @@ namespace {
 
 /// Inliner implementation that works with both, LPM (using SCC_t=CallGraph) and
 /// NPM (using SCC_t=LazyCallGraph::SCC)
-template <typename SCC_t> bool runScopInlinerImpl(Function *F, SCC_t &SCC) {
+template <typename SCC_t>
+bool runScopInlinerImpl(Function *F, SCC_t &SCC,
+                        IntrusiveRefCntPtr<vfs::FileSystem> FS) {
   // We do not try to inline non-trivial SCCs because this would lead to
   // "infinite" inlining if we are not careful.
   if (SCC.size() > 1)
@@ -50,7 +52,11 @@ template <typename SCC_t> bool runScopInlinerImpl(Function *F, SCC_t &SCC) {
     return false;
   }
 
-  PassBuilder PB;
+  PassBuilder PB(
+      /*TM=*/nullptr,
+      /*PipelineTuningOptions=*/{},
+      /*PGOOpt=*/{},
+      /*PIC=*/nullptr, std::move(FS));
   // Populate analysis managers and register Polly-specific analyses.
   LoopAnalysisManager LAM;
   FunctionAnalysisManager FAM;
@@ -96,7 +102,8 @@ template <typename SCC_t> bool runScopInlinerImpl(Function *F, SCC_t &SCC) {
 }
 } // namespace
 
-polly::ScopInlinerPass::ScopInlinerPass() {
+polly::ScopInlinerPass::ScopInlinerPass(IntrusiveRefCntPtr<vfs::FileSystem> FS)
+    : FS(std::move(FS)) {
   if (!polly::PollyAllowFullFunction) {
     report_fatal_error(
         "Aborting from ScopInliner because it only makes sense to run with "
@@ -113,6 +120,6 @@ PreservedAnalyses polly::ScopInlinerPass::run(llvm::LazyCallGraph::SCC &SCC,
                                               llvm::LazyCallGraph &CG,
                                               llvm::CGSCCUpdateResult &UR) {
   Function *F = &SCC.begin()->getFunction();
-  bool Changed = runScopInlinerImpl(F, SCC);
+  bool Changed = runScopInlinerImpl(F, SCC, FS);
   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
 }

polly/test/.clang-format

@@ -0,0 +1 @@
+ColumnLimit: 0

polly/test/CMakeLists.txt

@@ -3,6 +3,7 @@ set(LLVM_SHLIBEXT "${CMAKE_SHARED_MODULE_SUFFIX}")
 add_custom_target(check-polly)
 set_target_properties(check-polly PROPERTIES FOLDER "Polly/Meta")
 
+set(POLLY_TEST_C_COMPILER "")
 if(NOT LLVM_MAIN_SRC_DIR)
   find_program(LLVM_OPT NAMES opt HINTS ${LLVM_TOOLS_BINARY_DIR})
   find_program(LLVM_FILECHECK NAMES FileCheck HINTS ${LLVM_TOOLS_BINARY_DIR})
@@ -35,6 +36,11 @@ else ()
   set(LLVM_NOT "${LLVM_TOOLS_BINARY_DIR}/not")
   set(POLLY_TEST_EXTRA_PATHS "")
   set(POLLY_TEST_DEPS llvm-config opt LLVMPolly FileCheck not count)
+
+  if (TARGET clang)
+    set(POLLY_TEST_C_COMPILER "${LLVM_TOOLS_BINARY_DIR}/clang${CMAKE_EXECUTABLE_SUFFIX}")
+    list(APPEND POLLY_TEST_DEPS clang)
+  endif ()
 endif()
 
 if (POLLY_BUNDLED_ISL)

polly/test/CodeGen/RuntimeDebugBuilder/combine_different_values.c

@@ -1,23 +0,0 @@
-#define N 10
-void foo(float A[restrict], double B[restrict], char C[restrict],
-         int D[restrict], long E[restrict]) {
-  for (long i = 0; i < N; i++)
-    A[i] += B[i] + C[i] + D[i] + E[i];
-}
-
-int main() {
-  float A[N];
-  double B[N];
-  char C[N];
-  int D[N];
-  long E[N];
-
-  for (long i = 0; i < N; i++) {
-    __sync_synchronize();
-    A[i] = B[i] = C[i] = D[i] = E[i] = 42;
-  }
-
-  foo(A, B, C, D, E);
-
-  return A[8];
-}

polly/test/lit.cfg

@@ -4,11 +4,13 @@ import os
 import platform
 import re
 import subprocess
+import shlex
 
 import lit.formats
 import lit.util
 
 from lit.llvm import llvm_config
+from lit.llvm.subst import ToolSubst
 
 # Configuration file for the 'lit' test runner.
 
@@ -55,6 +57,25 @@ llvm_config.use_default_substitutions()
 tool_patterns = ['opt', 'polly-isl-test']
 llvm_config.add_tool_substitutions(tool_patterns)
 
+
+if config.polly_test_c_compiler:
+  config.suffixes += ['.c']
+  llvm_polly_link_into_tools = lit.util.pythonize_bool(config.llvm_polly_link_into_tools)
+  clang_args = []
+  if not llvm_polly_link_into_tools:
+    clang_args += [f'-fpass-plugin={config.polly_lib_dir}/LLVMPolly{config.llvm_shlibext}']
+  llvm_config.add_tool_substitutions([
+    ToolSubst("%clang",
+      command = config.polly_test_c_compiler,
+      extra_args = clang_args,
+      unresolved = "fatal"
+    )
+  ])
+  lit_config.note(f"Using Clang: {config.polly_test_c_compiler} {shlex.join(clang_args)}")
+else:
+  lit_config.note("Clang tests disabled: No compatible Clang found")
+
+
 # opt knows whether it is compiled with -DNDEBUG.
 import subprocess
 try:

polly/test/lit.site.cfg.in

@@ -4,12 +4,14 @@ config.llvm_src_root = "@LLVM_SOURCE_DIR@"
 config.llvm_obj_root = "@LLVM_BINARY_DIR@"
 config.llvm_tools_dir = lit_config.substitute("@LLVM_TOOLS_DIR@")
 config.llvm_libs_dir = lit_config.substitute("@LLVM_LIBS_DIR@")
+config.llvm_shlibext = "@LLVM_SHLIBEXT@"
 config.polly_obj_root = "@POLLY_BINARY_DIR@"
 config.polly_lib_dir = "@POLLY_LIB_DIR@"
 config.target_triple = "@LLVM_TARGET_TRIPLE@"
 config.llvm_polly_link_into_tools = "@LLVM_POLLY_LINK_INTO_TOOLS@"
 config.targets_to_build = "@TARGETS_TO_BUILD@"
 config.extra_paths = "@POLLY_TEST_EXTRA_PATHS@".split(";")
+config.polly_test_c_compiler = "@POLLY_TEST_C_COMPILER@"
 
 ## Check the current platform with regex
 import re

polly/test/polly.c

@@ -0,0 +1,9 @@
+// Sanity test for Polly in Clang
+// RUN: %clang %s -O2 -c -mllvm -polly -mllvm -polly-process-unprofitable -mllvm -print-pipeline-passes -o %t.o | FileCheck %s
+
+// CHECK: ,polly,
+
+void foo(int *A, int *B, int n) {
+  for (int i = 0; i < n; ++i)
+    A[i] += B[i];
+}