Support (optionally) emitting debuginfo w.r.t. LLVM source

This change adds two new environmental variables:
  - `JULIA_DUMP_IR` - when provided, this is a path that all emitted
    LLVM IR (post-optimization, just before machine code generation)
    will be saved to
  - `JULIA_DEBUGINFO` - when set to "LLVM-IR" this will run an additional
    pass on any emitted functions to rewrite their debuginfo to refer to
    the LLVM source, rather than the Julia source it was generated from

The `debugir` pass that rewrites the debuginfo is vendored from:
  https://github.com/vaivaswatha/debugir. For simplicity, this is just a
copy of the one file that we need for the pass.

Using both of these together allows `gdb` to open the dumped IR and means
you can step through LLVM IR line-by-line, print SSA values, etc. This can
be very useful for debugging segfaults, or issues in codegen.

THIRDPARTY.md

@@ -5,7 +5,7 @@ for exceptions.
 
 - [crc32c.c](https://stackoverflow.com/questions/17645167/implementing-sse-4-2s-crc32c-in-software) (CRC-32c checksum code by Mark Adler) [[ZLib](https://opensource.org/licenses/Zlib)].
 - [LDC](https://github.com/ldc-developers/ldc/blob/master/LICENSE) (for ccall/cfunction ABI definitions) [BSD-3]. The portion of code that Julia uses from LDC is [BSD-3] licensed.
-- [LLVM](https://releases.llvm.org/3.9.0/LICENSE.TXT) (for parts of src/disasm.cpp) [UIUC]
+- [LLVM](https://releases.llvm.org/3.9.0/LICENSE.TXT) (for parts of src/disasm.cpp and src/llvm-debugir.cpp) [UIUC]
 - [NetBSD](https://www.netbsd.org/about/redistribution.html) (for setjmp, longjmp, and strptime implementations on Windows) [BSD-3]
 - [Python](https://docs.python.org/3/license.html) (for strtod implementation on Windows) [PSF]
 - [FEMTOLISP](https://github.com/JeffBezanson/femtolisp) [BSD-3]

src/Makefile

@@ -52,7 +52,7 @@ RT_LLVMLINK :=
 CG_LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
-CODEGEN_SRCS := codegen jitlayers aotcompile debuginfo disasm llvm-simdloop \
+CODEGEN_SRCS := codegen jitlayers aotcompile debuginfo disasm llvm-simdloop llvm-debugir \
 	llvm-final-gc-lowering llvm-pass-helpers llvm-late-gc-lowering llvm-ptls \
 	llvm-lower-handlers llvm-gc-invariant-verifier llvm-propagate-addrspaces \
 	llvm-multiversioning llvm-alloc-opt llvm-alloc-helpers cgmemmgr llvm-remove-addrspaces \
@@ -327,6 +327,7 @@ $(build_shlibdir)/libllvmcalltest.$(SHLIB_EXT): $(SRCDIR)/llvm-codegen-shared.h
 $(BUILDDIR)/llvm-alloc-helpers.o $(BUILDDIR)/llvm-alloc-helpers.dbg.obj: $(SRCDIR)/llvm-codegen-shared.h $(SRCDIR)/llvm-pass-helpers.h $(SRCDIR)/llvm-alloc-helpers.h
 $(BUILDDIR)/llvm-alloc-opt.o $(BUILDDIR)/llvm-alloc-opt.dbg.obj: $(SRCDIR)/llvm-codegen-shared.h $(SRCDIR)/llvm-pass-helpers.h $(SRCDIR)/llvm-alloc-helpers.h
 $(BUILDDIR)/llvm-cpufeatures.o $(BUILDDIR)/llvm-cpufeatures.dbg.obj: $(SRCDIR)/jitlayers.h
+$(BUILDDIR)/llvm-debugir.o $(BUILDDIR)/llvm-debugir.dbg.obj: $(SRCDIR)/llvm-codegen-shared.h
 $(BUILDDIR)/llvm-demote-float16.o $(BUILDDIR)/llvm-demote-float16.dbg.obj: $(SRCDIR)/jitlayers.h
 $(BUILDDIR)/llvm-final-gc-lowering.o $(BUILDDIR)/llvm-final-gc-lowering.dbg.obj: $(SRCDIR)/llvm-gc-interface-passes.h
 $(BUILDDIR)/llvm-gc-invariant-verifier.o $(BUILDDIR)/llvm-gc-invariant-verifier.dbg.obj: $(SRCDIR)/llvm-codegen-shared.h

src/codegen.cpp

@@ -10395,6 +10395,24 @@ extern "C" void jl_init_llvm(void)
 #endif
 #endif
 
+    const char *debuginfo_mode = getenv("JULIA_DEBUGINFO"); // JULIA_DEBUGINFO="LLVM-IR" JULIA_DUMP_IR="path/to/dir"
+    if (debuginfo_mode) {
+        if (strcasecmp(debuginfo_mode, "julia-source") == 0)
+            jl_ExecutionEngine->get_debuginfo_mode() = jl_debuginfo_emission_mode_t::julia_source;
+        else if (strcasecmp(debuginfo_mode, "llvm-ir") == 0)
+            jl_ExecutionEngine->get_debuginfo_mode() = jl_debuginfo_emission_mode_t::llvm_ir;
+        else if (strcmp(debuginfo_mode, "") != 0)
+            fprintf(stderr, "warning: unexpected argument to 'JULIA_DEBUGINFO' env var: \"%s\"\n", debuginfo_mode);
+    }
+
+    const char *dump_debugir_directory = getenv("JULIA_DUMP_IR");
+    if (dump_debugir_directory && strcmp(dump_debugir_directory, "") != 0) {
+        llvm::SmallString<PATH_MAX> AbsoluteFileName{};
+        llvm::sys::fs::expand_tilde(Twine(dump_debugir_directory), AbsoluteFileName);
+        llvm::sys::fs::make_absolute(AbsoluteFileName);
+        jl_ExecutionEngine->get_dump_debugir_directory() = AbsoluteFileName.str();
+    }
+
     cl::PrintOptionValues();
 }
 

src/jitlayers.cpp

@@ -1314,6 +1314,39 @@ static void registerRTDyldJITObject(orc::MaterializationResponsibility &MR,
 #endif
 
 namespace {
+
+    static std::string createDebugIRName(Module const &M) {
+        std::string path = jl_ExecutionEngine->get_dump_debugir_directory();
+        if (!path.empty()) {
+            path += llvm::sys::path::get_separator();
+        }
+
+        std::string filename{};
+        for (auto &F : M.functions()) {
+            if (F.isDeclaration() || F.getName().starts_with("jfptr_"))
+                continue;
+
+            // Sanitize the filename by allowing only "[a-zA-Z0-9_\-\.]*"
+            filename = F.getName().str();
+            std::replace_if(std::begin(filename), std::end(filename),[](const char &ch) {
+                return !(std::isalnum(ch) || ch == '_' || ch == '-' || ch == '.');
+            }, '_');
+
+            break;
+        }
+
+        // If we couldn't find a name to use, just use some unique integer
+        if (filename.empty()) {
+            static _Atomic(uint64_t) counter{1};
+            filename = std::to_string(jl_atomic_fetch_add_relaxed(&counter, 1));
+        }
+
+        path += filename;
+        path += ".ll";
+
+        return path;
+    }
+
     static std::unique_ptr<TargetMachine> createTargetMachine() JL_NOTSAFEPOINT {
         TargetOptions options = TargetOptions();
 
@@ -1509,6 +1542,44 @@ namespace {
                     JL_TIMING(LLVM_JIT, JIT_Opt);
                     //Run the optimization
                     (****PMs[PoolIdx]).run(M);
+
+                    bool debug_ir = jl_ExecutionEngine->get_debuginfo_mode() == jl_debuginfo_emission_mode_t::llvm_ir;
+                    bool dump_ir = !jl_ExecutionEngine->get_dump_debugir_directory().empty();
+                    if (!M.functions().empty() && (debug_ir || dump_ir)) {
+
+                        // Generate a debug filename for the emitted IR
+                        std::string debug_name = std::move(createDebugIRName(M));
+
+                        // If requested, rewrite all debuginfo to reference the LLVM IR itself
+                        std::unique_ptr<Module> displayM;
+                        if (debug_ir) {
+                            // displayM is the debug-stripped 'source' that the debuginfo now refers to
+                            displayM = debugir::createDebugInfo(M, "", debug_name);
+                        }
+
+                        // Emit the IR that was compiled
+                        if (dump_ir) {
+                            std::error_code EC;
+                            raw_fd_ostream OS_dbg(debug_name, EC, sys::fs::OF_Text);
+                            if (displayM) {
+                                displayM->print(OS_dbg, nullptr);
+                            } else {
+                                M.print(OS_dbg, nullptr);
+                            }
+
+                            // Emit the "instrumented" IR (unneeded unless you are debugging the debuginfo
+                            // or running the instrumented IR in isolation)
+                            if (0 && displayM) {
+                                // Replace ".ll" suffix with ".dbg.ll"
+                                debug_name.resize(debug_name.size() - 3);
+                                debug_name += ".dbg.ll";
+                                std::error_code EC;
+                                raw_fd_ostream OS_dbg(debug_name, EC, sys::fs::OF_Text);
+                                M.print(OS_dbg, nullptr);
+                            }
+                        }
+                    }
+
                     assert(!verifyLLVMIR(M));
                 }
 
@@ -1898,6 +1969,7 @@ JuliaOJIT::JuliaOJIT()
     JD(ES.createBareJITDylib("JuliaOJIT")),
     ExternalJD(ES.createBareJITDylib("JuliaExternal")),
     DLSymOpt(std::make_unique<DLSymOptimizer>(false)),
+    debuginfo_mode(jl_debuginfo_emission_mode_t::julia_source),
 #ifdef JL_USE_JITLINK
     MemMgr(createJITLinkMemoryManager()),
     ObjectLayer(ES, *MemMgr),

src/jitlayers.h

@@ -367,6 +367,28 @@ using CompilerResultT = Expected<std::unique_ptr<llvm::MemoryBuffer>>;
 using OptimizerResultT = Expected<orc::ThreadSafeModule>;
 using SharedBytesT = StringSet<MaxAlignedAllocImpl<sizeof(StringSet<>::MapEntryTy)>>;
 
+enum class jl_debuginfo_emission_mode_t {
+
+    // Source-referenced debuginfo (standard behavior)
+    //
+    // Preserves and emits any debuginfo in the IR from Julia source.
+    julia_source = 0,
+
+    // Julia IR-referenced debuginfo
+    //
+    // Replaces all debuginfo with references to the Julia SSAIR itself (treating the IR code as
+    // the 'program source'). Emits a text copy of all emitted IR to 'dump_debugir_directory' so
+    // the IR source is available when using a debugger.
+    /* julia_ir, not supported (yet) */
+
+    // LLVM IR-referenced debuginfo
+    //
+    // Replaces all debuginfo with references to the LLVM IR itself (treating the LLVM IR as the
+    // 'program source'). Emits a text copy of all emitted IR to 'dump_debugir_directory' so the
+    // IR source is available when using a debugger.
+    llvm_ir,
+};
+
 class JuliaOJIT {
 private:
     // any verification the user wants to do when adding an OwningResource to the pool
@@ -582,9 +604,16 @@ class JuliaOJIT {
     jl_locked_stream &get_dump_llvm_opt_stream() JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER {
         return dump_llvm_opt_stream;
     }
+    std::string &get_dump_debugir_directory() JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER {
+        return dump_debugir_directory;
+    }
     std::string getMangledName(StringRef Name) JL_NOTSAFEPOINT;
     std::string getMangledName(const GlobalValue *GV) JL_NOTSAFEPOINT;
 
+    jl_debuginfo_emission_mode_t &get_debuginfo_mode() JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER {
+        return debuginfo_mode;
+    }
+
     // Note that this is a potential safepoint due to jl_get_library_ and jl_dlsym calls
     // but may be called from inside safe-regions due to jit compilation locks
     void optimizeDLSyms(Module &M) JL_NOTSAFEPOINT_LEAVE JL_NOTSAFEPOINT_ENTER;
@@ -610,6 +639,9 @@ class JuliaOJIT {
     jl_locked_stream dump_emitted_mi_name_stream;
     jl_locked_stream dump_compiles_stream;
     jl_locked_stream dump_llvm_opt_stream;
+    std::string      dump_debugir_directory;
+
+    jl_debuginfo_emission_mode_t debuginfo_mode;
 
     std::mutex llvm_printing_mutex{};
     SmallVector<std::function<void()>, 0> PrintLLVMTimers;

src/llvm-codegen-shared.h

@@ -527,3 +527,12 @@ void ConstantUses<U>::forward()
     }
 }
 }
+
+namespace debugir {
+
+// Attaches debug info to M, assuming it is parsed from Directory/Filename.
+// Returns a module for display in debugger devoid of any debug info.
+std::unique_ptr<llvm::Module>
+createDebugInfo(llvm::Module &M, std::string Directory, std::string Filename);
+
+} // namespace debugir