PR19091, libcxxabi fix of wrong demangling of "operator>(int)"

[rsmith]

Summary

Problem is in cxa_demangle.cpp, "parse_encoding" method.

Method checks previously demangled characters to detect whether template arguments were parsed, and whether it can proceed to type parsing.
So, method misrecognises ">" as template closing tag. It wouldn't be good if we check that closing tag is not a tail of "operator>", since it won't solve whole problem. It still will conflict with constructions like "<despicable_operator>".
Current patch fixes this problem by adding extra space into demangling sequence for GT operator: "operator> ".
Though generally, it would be good if we will keep parsing stack information somewhere.

[mclow.lists]

I'm not sure that this is a good solution; I don't think that there's any other place do we add a space in the demangled name (at least between things that the compiler would accept w/o the space)

[dyatkovskiy]

I have the same opinion. IMHO, the best solution is to extend Db with information about what was parsed. But how we could solve it now? Is there any other ways to distinguish operator> from tail of "<some_operator>" ?

[dyatkovskiy]

So guys, how do you find this fix?

[rsmith]

This does not seem like a reasonable approach. The problem is that we're looking at the contents of the already-demangled buffer to determine how to demangle and then making incorrect assumptions about it. We should stop making those assumptions.

It doesn't seem difficult for parse_name to track whether the last component in the <name> within an <encoding> had template arguments, and that's what we need to know to determine how to demangle a function or function template specialization.

[rsmith]

Fixed in r208611.

src/cxa_demangle.cpp

         case 'g':
             switch (first[1])
             {
             case 'e':
                 db.names.push_back("operator>=");
                 first += 2;
                 break;
             case 't':
-                db.names.push_back("operator>");
+                // Add extra space and avoid collisions with templates.
+                db.names.push_back("operator> ");
                 first += 2;
                 break;
             }
             break;
         case 'i':
             if (first[1] == 'x')
             {
                 db.names.push_back("operator[]");

test/test_demangle.cpp

 
 const char* cases[][2] =
 {
     {"_Z1A", "A"},
     {"_Z1Av", "A()"},
     {"_Z1A1B1C", "A(B, C)"},
     {"_Z4testI1A1BE1Cv", "C test<A, B>()"},
     {"_Z4testI1A1BET0_T_S3_", "B test<A, B>(A, A)"},
+    {"_ZN1SgtEi", "S::operator> (int)"},
     {"_ZN13dyldbootstrap5startEPK12macho_headeriPPKcl", "dyldbootstrap::start(macho_header const*, int, char const**, long)"},
     {"_ZN4dyld17getExecutablePathEv", "dyld::getExecutablePath()"},
     {"_ZN4dyld22mainExecutablePreboundEv", "dyld::mainExecutablePrebound()"},
     {"_ZN4dyld14mainExecutableEv", "dyld::mainExecutable()"},
     {"_ZN4dyld21findImageByMachHeaderEPK11mach_header", "dyld::findImageByMachHeader(mach_header const*)"},
     {"_ZN4dyld26findImageContainingAddressEPKv", "dyld::findImageContainingAddress(void const*)"},
     {"_ZN4dyld17clearErrorMessageEv", "dyld::clearErrorMessage()"},
     {"_ZN4dyld15getErrorMessageEv", "dyld::getErrorMessage()"},