Some g++ symbols aren't demangled (.localalias & .cold)

[blackgnezdo]

I linked glog 0.6.0 using this NixOS recipe (with gflags dependency removed) into my c++ application. I see a stack trace with some C++ symbols properly demangled while a few remain mangled (and good enough for c++filt to decode). E.g.

    @     0x7f051d8a618a _ZN4mlir6detail17OpToOpPassAdaptor3runEPNS_4PassEPNS_9OperationENS_15AnalysisManagerEbj.localalias
    @     0x7f051d8a6599 _ZN4mlir6detail17OpToOpPassAdaptor11runPipelineERNS_13OpPassManagerEPNS_9OperationENS_15AnalysisManagerEbjPNS_16PassInstrumentorEPKNS_19PassInstrumentation18PipelineParentInfoE.localalias
    @     0x7f051d8a54ef _ZN4mlir6detail17OpToOpPassAdaptor23runOnOperationAsyncImplEb.localalias
    @     0x7f051d8a5fd4 _ZN4mlir6detail17OpToOpPassAdaptor3runEPNS_4PassEPNS_9OperationENS_15AnalysisManagerEbj.localalias
    @     0x7f051d8a6599 _ZN4mlir6detail17OpToOpPassAdaptor11runPipelineERNS_13OpPassManagerEPNS_9OperationENS_15AnalysisManagerEbjPNS_16PassInstrumentorEPKNS_19PassInstrumentation18PipelineParentInfoE.localalias
    @     0x7f051d8a7420 mlir::PassManager::run()
    @           0x525cf8 main
    @     0x7f05088e2790 __libc_start_main
    @           0x5c3e7a _start

which after c++filt become a pleasant to read

    @     0x7f051d8a618a mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) [clone .localalias]
    @     0x7f051d8a6599 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) [clone .localalias]
    @     0x7f051d8a54ef mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool) [clone .localalias]
    @     0x7f051d8a5fd4 mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) [clone .localalias]
    @     0x7f051d8a6599 mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) [clone .localalias]
    @     0x7f051d8a7420 mlir::PassManager::run()
    @           0x525cf8 main
    @     0x7f05088e2790 __libc_start_main
    @           0x5c3e7a _start

I activate the stack trace with google::InstallFailureSignalHandler(); no further flags or configuration.

Is there some kind of limit on the length of mangled identifiers or some other option I'm missing?

[blackgnezdo]

I played a bit with demangle unit test and looks like the failure is due to the missing .localalias support. This test shows the missing functionality:

diff --git a/src/demangle_unittest.cc b/src/demangle_unittest.cc
index ddc90b0..d3d8b52 100644
--- a/src/demangle_unittest.cc
+++ b/src/demangle_unittest.cc
@@ -121,6 +121,8 @@ TEST(Demangle, Clones) {
   EXPECT_FALSE(Demangle("_ZL3Foov.clone.foo", tmp, sizeof(tmp)));
   // Invalid (.constprop. not followed by number), should not demangle.
   EXPECT_FALSE(Demangle("_ZL3Foov.isra.2.constprop.", tmp, sizeof(tmp)));
+  char large[256];
+  EXPECT_TRUE(Demangle("_ZN4mlir6detail17OpToOpPassAdaptor3runEPNS_4PassEPNS_9OperationENS_15AnalysisManagerEbj.localalias", large, sizeof(large)));
 }
 
 TEST(Demangle, FromFile) {

The test passes if I e.g. remove .localalias suffix.

[blackgnezdo]

A very naive way to make the test pass would be to add

diff --git a/src/demangle.cc b/src/demangle.cc
index 9276c5b..528a4ee 100644
--- a/src/demangle.cc
+++ b/src/demangle.cc
@@ -303,6 +303,7 @@ static bool IsDigit(char c) {
 // a function clone suffix.
 static bool IsFunctionCloneSuffix(const char *str) {
   size_t i = 0;
+  if (strcmp(str, ".localalias") == 0) return true;
   while (str[i] != '\0') {
     // Consume a single .<alpha>+.<digit>+ sequence.
     if (str[i] != '.' || !IsAlpha(str[i + 1])) {

[blackgnezdo]

I noticed another similar case, this time it's a different suffix: _ZN9__gnu_cxx27__verbose_terminate_handlerEv.cold. Looks like a more complete fix would be to allow mostly arbitrary alpha+ suffixes without requiring they are followed by number+.

[sergiud]

The glog native demangler is pretty much incomplete and should be regarded as obsolete.

glog 0.7.0 will introduce the use of GCC/LLVM provided ABI demangler which will solve als these issues.

[sergiud]

#1023 has landed in master. Please give it a try.