gdb bug list John Gilmore, gnu@cygnus.com This bug list is probably not up to date or accurate, but it reflects some known bugs in gdb, if you are into bug-hunting. It should be possible to use symbols from shared libraries before we know exactly where the libraries will be loaded. E.g. "b perror" before running the program. This could maybe be done as an extension of the "breakpoint re-evaluation" after new symbols are loaded. Make single_step() insert and remove breakpoints in one operation. Speed up single stepping by avoiding extraneous ptrace calls. Speed up single stepping by not inserting and removing breakpoints each time the inferior starts and stops. Breakpoints should not be inserted and deleted all the time. Only the one(s) there should be removed when we have to step over one. Support breakpoints that don't have to be removed to step over them. Speed up watchpoints by not single-stepping them, but do something faster like single-line execution. Speed them up tremendously on machines that have watchpoint registers. Update gdbint.texinfo to include doc on the directory structure and the various tricks of building gdb. Do a tutorial in gdb.texinfo on how to do simple things in gdb. E.g. how to set a breakpoint that just prints something and continues. How to break on aborts. Etc. Provide "voodoo" debugging of core files. This creates a zombie process as a child of the debugger, and loads it up with the data, stack, and regs of the core file. This allows you to call functions in the executable, to manipulate the data in the core file. GDB reopens the source file on every line, as you "next" through it. Referencing the vtbl member of a struct doesn't work. It prints OK if you print the struct, but it gets 0 if you try to deref it. Persistent command history: A feature where you could save off a list of the commands you did, so you can edit it into something that will bring the target to the same place every time you source it. This would also be useful for automated fast watchpointing; if you go past the place where it watchpoints, you just start it over again and do it more carefully. Deal with the SunOS 4.0 and 4.1.1 ptrace bug that loses the registers if the stack is paged out. Finish the C++ exception handling stub routines. Lint points them out as unused statics functions. Perhaps "i source" should take an argument like that of "list". See if coredep.c's fetch_core_registers can be used on more machines. E.g. MIPS (mips-xdep.c). unpack_double() does not handle IEEE float on the target unless the host is also IEEE. Death on a vax. Test cross-debugging Unix-to-Unix. Integrate the RAPP remote debugging protocol. `rapp' runs under unix and acts as a remote stub (like rem-multi.shar distributed with GDB version 3). Currently it just works over UDP (network), not over a serial line. To get it running * Compile GDB on the host machine as usual * Compile rapp on the target machine, giving for both host and target the type of the target machine * Install "gdb" in /etc/services on both machines. Set up interface between GDB and INFO so that you can hop into interactive INFO and back out again. When running under Emacs, should use Emacs info, else fork the info program. Installation of GDB should install its texinfo files into the info tree automagically, including the readline texinfo files. "help address" ought to find the "help set print address" entry. Remove the VTBL internal guts from printouts of C++ structs, unless vtblprint is set. Remove "at 0xnnnn" from the "b foo" response, if `print address off' and if it matches the source line indicated. The prompt at end of screen should accept space as well as CR. "List" should put you into a pseudo-"more" where you can hit space to get more, forever to eof. Check STORE_RETURN_VALUE on all architectures. Check near it in tm-sparc.h for other bogosities. Check for storage leaks in GDB, I'm sure there are a lot! vtblprint of a vtbl should demangle the names it's printing. Backtrace should point out what the currently selected frame is, in its display, perhaps showing "@3 foo (bar, ...)" rather than "#3 foo (bar, ...)". "i program" should work for core files, and display more info, like what actually caused it to die. Hitting ^Z to an inferior doesn't work right, it takes several continues to make it actually go. "x/10i" should shorten the long name, if any, on subsequent lines. Check through the code for FIXME comments and fix them. dbxread.c, blockframe.c, and plenty more. "next" over a function that longjumps, never stops until next time you happen to get to that spot by accident. E.g. "n" over execute_command which has an error. Watchpoints seem not entirely reliable, though they haven't failed me recently. "set zeroprint off", don't bother printing members of structs which are entirely zero. Useful for those big structs with few useful members. GDB does four ioctl's for every command, probably switching terminal modes to/from inferior or for readline or something. terminal_ours versus terminal_inferior: cache state. Switch should be a noop if the state is the same, too. ptype $i6 = void??! Clean up invalid_float handling so gdb doesn't coredump when it tries to access a NaN. While this might work on SPARC, other machines are not configured right. "b value_at ; commands ; continue ; end" stops EVERY OTHER TIME! Then once you enter a command, it does the command, runs two more times, and then stops again! Bizarre... (This behaviour has been modified, but it is not yet 100% predictable when e.g. the commands call functions in the child, and while there, the child is interrupted with a signal, or hits a breakpoint.) Symbol completion with TAB does not unmangle names! help completion, help history should work. Symbol completion doesn't handle e.g. W::f. (symtab.c, make_symbol_completion_list). Check that we can handle stack trace through varargs AND alloca in same function, on 29K. wait_for_inferior loops forever if wait() gives it an error. "i frame" shows wrong "arglist at" location, doesn't show where the args should be found, only their actual values. Symbolic display of addrs, (& disassembly prefixes), don't show static fns, e.g. enable_command in gdb. 'ptype yylval' ==> "union YYSTYPE { ..... }". However, it is not a union YYSTYPE, but is simply a YYSTYPE, which is a typedef for an unnamed union. There should be a way for "set" commands to validate the new setting before it takes effect. The "display" command should become the "always" command, e.g. "always print XXX" "always p/xxx XXX" "always echo foo" "always call XXX" "always x/i $pc", etc. A mess of floating point opcodes are missing from sparc-opcode.h. Also, a little program should test the table for bits that are overspecified or underspecified. E.g. if the must-be-ones bits and the must-be-zeroes bits leave some fields unexamined, and the format string leaves them unprinted, then point this out. If multiple non-alias patterns match, point this out too. Finally, there should be a sparc-optest.s file that tries each pattern out. This file should end up coming back the same (modulo transformation comments) if fed to "gas" then the .o is fed to gdb for disassembly. Merge the xxx-opcode.h files with gas again... Eliminate all the core_file_command's in all the xdep files. Eliminate separate declarations of registers[] everywhere. "ena d" is ambiguous, why? "ena delete" seems to think it is a command! Line numbers are off in some spots. In proceed() at 1st "oneproc = 1", it seems to run that statement, but it doesn't actually. Perhaps move the tdep and xdep files, and the tm and xm files, into a config subdirectory. If not, at least straighten out their names so that they all start with the machine name. inferior_status should include stop_print_frame. It won't need to be reset in wait_for_inferior after bpstat_stop_status call, then. i line VAR produces "Line number not known for symbol ``var''.". I thought we were stashing that info now! Make sure we can handle executables with no symbol info, e.g. /bin/csh. We should be able to write to random files at hex offsets like adb. Make "target xxx" command interruptible. Handle add_file with separate text, data, and bss addresses. Maybe handle separate addresses for each segment in the object file? Handle free_named_symtab to cope with multiply-loaded object files in a dynamic linking environment. Should remember the last copy loaded, but not get too snowed if it finds references to the older copy. The original BFD core dump reading routine would itself coredump when fed a garbage file as a core file. Does the current one? Generalize and Standardize the RPC interface to a target program, improve it beyond the "ptrace" interface, and see if it can become a standard for remote debugging. Remove all references to: text_offset data_offset text_data_start text_end exec_data_offset ... now that we have BFD. All remaining are in machine dependent files. When quitting with a running program, if a core file was previously examined, you get "Couldn't read float regs from core file"...if indeed it can't. generic_mourn_inferior... Check signal argument to remote proceed's and error if set. Sort help and info output. Re-organize help categories into things that tend to fit on a screen and hang together. renote-nindy.c handles interrupts poorly; it error()s out of badly chosen places, e.g. leaving current_frame zero, which causes core dumps on the next command. Add in commands like ADB's for searching for patterns, etc. We should be able to examine and patch raw unsymboled binaries as well in gdb as we can in adb. (E.g. increase the timeout in /bin/login without source). Those xdep files that call register_addr without defining it are probably simply broken. When reconfiguring this part of gdb, I could only make guesses about how to redo some of those files, and I probably guessed wrong, or left them "for later" when I have a machine that can attempt to build them. Use the complain() mechanism for handling all the error() calls in dbxread.c, and in similar situations in coffread.c and mipsread.c. When doing "step" or "next", if a few lines of source are skipped between the previous line and the current one, print those lines, not just the last line of a multiline statement. When searching for C++ superclasses in value_cast in valops.c, we must not search the "fields", only the "superclasses". There might be a struct with a field name that matches the superclass name. This can happen when the struct was defined before the superclass (before the name became a typedef). Handling of "&" address-of operator needs some serious overhaul for ANSI C and consistency on arrays and functions. For "float point[15];": ptype &point[4] ==> Attempt to take address of non-lvalue. For "char *malloc();": ptype malloc ==> "char *()"; should be same as ptype &malloc ==> "char *(*)()" call printf ("%x\n", malloc) ==> wierd value, should be same as call printf ("%x\n", &malloc) ==> correct value Fix dbxread.c symbol reading in the presence of interrupts. It currently leaves a cleanup to blow away the entire symbol table when a QUIT occurs. Mipsread.c reads include files depth-first, because the dependencies in the psymtabs are way too inclusive (it seems to me). Figure out what really depends on what, to avoid recursing 20 or 30 times while reading real symtabs. value_add() should be subtracting the lower bound of arrays, if known, and possibly checking against the upper bound for error reporting. mipsread.c symbol table allocation and deallocation should be checked. My suspicion is that it's full of memory leaks. SunOS should have a target_lookup_symbol() for common'd things allocated by the shared library linker ld.so. When listing source lines, check for a preceding \n, to verify that the file hasn't changed out from under us. When listing source lines, eat leading whitespace corresponding to the line-number prefix we print. This avoids long lines wrapping. mipsread.c needs to check for old symtabs and psymtabs for the same files, the way it happens for dbxread.c and coffread.c, for VxWorks incremental symbol table reloading. When attached to a non-child process, ^C or other signals are not propagated to the child. Do this in the GDB signal handler, using target_kill(). AMD version: ^C should do ^Ak to stop ebmon.