Debugging GHC-compiled code with gdb

If you're working on a GHC port, tweaking the runtime, or modifying the code generator, at some point you will have to become familiar with how to use gdb for debugging crashes in GHC-compiled programs.

Debugging of Haskell code in gdb got better recently, as you can ask GHC to generate DWARF annotations in the compiled Haskell code, which gdb can use to provide useful backtraces in Haskell code and to map symbol names to Haskell source code location. To get this, compile with the -g option. (for the best results, use GHC 8.4.1 or later).

The DWARF annotations can be a bit flaky still, and then we might be left with having to debug the compiled Haskell code at the assembly level. This can seem quite scary, but there are lots of resources on this wiki to help you understand what's going on.

The recipe

Usually you want to do something along these lines:

Compile all the Haskell code (including libraries) -g to get DWARF annotations.
Compile with -dcore-lint to make sure the crash isn't caused by the compiler generating incorrect code in a way that can be detected statically.
Compile with -debug! This links the program with a debugging version of the runtime, that includes extra assertions, debugging code, and debugging info for gdb (if you compile using cabal, use the unstripped binary from dist/build/).
Don't compile with -threaded (assuming the bug still happens with the non-threaded runtime).
Use runtime debugging options to help narrow down the fault (see also the relevant User Manual section).
Grab our gdb macros.
Run the program in gdb until it crashes, type where to find out whether the crash is in the RTS or in Haskell code. Hopefully you'll get a useful backtrace through both Haskell and C code, and if you compiled with -g you should see Haskell source code locations. If the crash was in the runtime, then you can use gdb's full C debugging facilities to track down the problem.
Make the crash happen as early as possible. -debug already turned on lots of assertions, one of which might trigger. Also, try +RTS -DS which turns on a bunch of extra "sanity checking", i.e. expensive assertions about the state of the runtime at regular points. One thing this enables is a full sweep of the heap after each garbage collection to make sure there are no dangling references. Also, it fills all free memory with the value 0xaaaaaaaa (the sound of GHC disappearing down a hole).
Look at the fragment of code that crashed: disassemble. It should be pretty small (code blocks in GHC-generated code usually are). Look at the instruction that caused the crash - the address is given in the where output earlier. If you can tell what the problem is right there, then great.
Poke around in the heap and stack and see if you can figure out what has gone wrong. See "poking around", below.
Make the crash happen repeatably, if possible. That means after the exact same number of garbage collections and context switches each time. Use +RTS -C0 to turn off timer-based context switches. Setting +RTS -V0 disables the RTS timer entirely. Disabling the RTS timer altogether can make the point at which the program crashes reproducible, and is very helpful when using the "going back in time" method described below. If you're debugging heap profiling, add -i0 too.
Similarly, if possible, turn off address space layout randomization (ASLR). On Linux you can do this with "echo 0 > /proc/sys/kernel/randomize_va_space".
If you still haven't figured it out, you probably need to go back in time a bit to find out what was happening just before the crash. This is the tricky bit, because going back in time isn't directly supported by gdb, we need to find a way to hit a breakpoint a short time before the crash happened. See "going back in time", below.

Debugging stack overflows

Set a breakpoint on threadStackOverflow, and when it gets hit p *tso.stackobj will tell you the stack pointer, e.g. sp = 0x7ffff69fe228. Then pmem 0x7ffff69fe228 64 will show you what's at the top of the stack.

Going back in time

There's a great tool called rr for reverse debugging. Try it out - if you're lucky it will work and you'll be able to single-step backwards and set breakpoints and watchpoints in the past. If rr doesn't work for you for whatever reason, try the following tricks.

Set a breakpoint on a code fragment that is regularly executed. Some good examples are stg_upd_frame_info (the standard update code), stg_upd_frame_1_info (the update code for a 1-tagged object, eg. a cons cell), stg_ap_p_info (the apply code for a single pointer argument).
- break stg_upd_frame_info
Ignore that breakpoint for ever
- ignore 1 9999999
Run the program
- run
Find out how many times the breakpoint was hit
- info break
Suppose it was hit 4325 times, then next time we'll ignore it for 4324 times (i.e. subtract one), which will stop on the 4325th time, just before the crash.
- ignore 1 4324
Run the program
- run
Single step until the crash happens
- si...
If the crash doesn't happen for a long time, try picking another break point (e.g. something you stepped through this time). You might have to repeat this a few times to get to a useful point.
Sometimes the crash is a long way into the program, and setting a frequently accessed breakopint like stg_upd_frame_info means the program just runs too slowly; in this case you might have to use a 2-stage strategy instead. Set a breakpoint on a less frequently accessed point (a good one I use for this is GarbageCollect, which stops at each GC), do the ignore thing on this, and then switch to a more frequently accessed breakpoint when you are closer to the crash.

Mapping symbol names to source code

Gdb's list command is useful for this, if the code was compiled with -g:

(gdb) list *s5aO_info
0x7ffff5762138 is in s5aO_info (compiler/GHC/Types/Name/Cache.hs:95).
90	        Just occ_env -> lookupOccEnv occ_env occ
91	
92	extendOrigNameCache :: OrigNameCache -> Name -> OrigNameCache
93	extendOrigNameCache nc name
94	  = ASSERT2( isExternalName name, ppr name )
95	    extendNameCache nc (nameModule name) (nameOccName name) name
96	
97	extendNameCache :: OrigNameCache -> Module -> OccName -> Name -> OrigNameCache
98	extendNameCache nc mod occ name
99	  = extendModuleEnvWith combine nc mod (unitOccEnv occ name)

The list command can also be used with a raw address, like list *0x7ffff5762138. This is often better, because a symbol name can be ambiguous, whereas an address is unique.

To go in the opposite direction, use info line:

(gdb) info line compiler/GHC/Types/Name/Cache.hs:95
Line 95 of "compiler/GHC/Types/Name/Cache.hs"
   starts at address 0x7ffff5762138 <s5aO_info>
   and ends at 0x7ffff5762141 <s5aO_info+9>.

Decoding symbol names

Symbols in GHC are encoded using something called the Z-encoding (see compiler/GHC/Utils/Encoding.hs). Basically special symbols are replaced by sequences beginning with z or Z. eg. state# becomes statezh. The letter z itself is replaced by zz.

External symbols follow the pattern moduleidentifierkind. Where module is the module name, identifier is the identifier in the STG code that this symbol relates to, and kind is the kind of symbol:

`closure`	a static closure
`info`	an info table, also entry code or return address
`con_info`	an info table for a constructor
`fast`	a primitive

(see compiler/GHC/Cmm/CLabel.hs for a table of these). Note that if you're matching up assembly with C-- and (info) tables next to code is enabled (as it is by default), then code that is named entry is equivalent to info symbols in the final binary.

For example:

GHCziBase_ZMZN_closure

Means the static closure for [] in module GHC.Base.

DataziList_foldlzq_info

Means the entry code (or info table) for Data.List.foldl'.

Internal symbols are of the form identifier_kind, i.e. the same as external symbols but without the module name.

Poking around

There are two things you usually want to inspect inside a code fragment: the contents of heap objects, and the contents of the stack. A full description of the layout of these things is here: Commentary/Rts/Storage; what follows is a basic introduction to looking at heap & stack objects in a running program.

You can display memory in gdb with something like x/4a to display 4 words of memory, or using our gdb macros you get slightly nicer output:

(gdb) p4 $rbx
0x2b35cc6c99a0: 0x4482d8 <stg_IND_1_info>
0x2b35cc6c9998: 0x2b35cc6c9978
0x2b35cc6c9990: 0x5
0x2b35cc6c9988: 0x404e30 <GHCziBase_Izh_con_info>

p4 displays 4 words of memory backwards, i.e. lower addresses at the bottom. In this case I'm displaying memory pointed to by the register rbx, which corresponds to the STG register R1 on a recent x86_64 build. Check includes/stg/MachRegs.h to see which machine registers correspond to which STG registers on your platform.

In the example above, I can see that $rbx points to a heap closure for the Int value 5. Closures always consist of an info pointer (GHCziBase_Izh_con_info in this case, the I# constructor), followed by any number of payload words (just one word containing the value 5, here). Full details on closure layouts are in includes/rts/storage/Closures.h.

It looks like the next word contains garbage, probably because it is the next free location in the heap.

We can see the info tables using the pinfo macro:

(gdb) pinfo &GHCziBase_Izh_con_info
$2 = {layout = {payload = {ptrs = 0, nptrs = 1}, bitmap = 4294967296, 
    large_bitmap_offset = 0, __pad_large_bitmap_offset = 0, 
    selector_offset = 4294967296}, type = 3, srt_bitmap = 0, 
  code = 0x404e30 "ÿe"}

Info tables tell the garbage collector and other parts of the system about the layout of closures, they are rarely used during actual execution. The info pointer of a closure actually points to the entry code for the closure; the info table lives just before the entry code (this is a trick used by GHC so that the common operation of jumping to the entry code for a closure can be done with a single indirection). The layout of info tables is defined in includes/rts/storage/InfoTables.h.

To display the stack, you need to know what the Sp register is mapped to on this platform. On x86_64 you'll find it in $rbp:

(gdb) p8 $rbp
0x2ae9697ccc00: 0x1
0x2ae9697ccbf8: 0x40b8c8 <GHCziBase_plusInt_info+32>
0x2ae9697ccbf0: 0x2ae9696c9868
0x2ae9697ccbe8: 0x44a2d8 <stg_upd_frame_info>
0x2ae9697ccbe0: 0x1
0x2ae9697ccbd8: 0x40b8c8 <GHCziBase_plusInt_info+32>
0x2ae9697ccbd0: 0x2ae9696c98a8
0x2ae9697ccbc8: 0x44a2d8 <stg_upd_frame_info>

It looks like the stack has an update frame on the top, followed by some stack frame from GHC.Base, followed by another update frame.

Return addresses on the stack have info tables just like heap closures, and you can display them with pinfo in the same way:

(gdb) pinfo 0x40b8c8
$4 = {layout = {payload = {ptrs = 65, nptrs = 0}, bitmap = 65, 
    large_bitmap_offset = 65, __pad_large_bitmap_offset = 65, 
    selector_offset = 65}, type = 36, srt_bitmap = 0, 
  code = 0x40b8c8 "I\203Ä\020M9üw$H\213E\bI\003E\bIÇD$ø0N@"}

However, return addresses have special info tables with more information; to display the whole thing, use prinfo

(gdb) prinfo 0x40b8c8
$5 = {srt_offset = 4241688, __pad_srt_offset = 6684481, i = {layout = {
      payload = {ptrs = 65, nptrs = 0}, bitmap = 65, large_bitmap_offset = 65, 
      __pad_large_bitmap_offset = 65, selector_offset = 65}, type = 36, 
    srt_bitmap = 0, code = 0x40b8c8 "I\203Ä\020M9üw$H\213E\bI\003E\bIÇD$ø0N@"}}

The type field tells us what kind of object this is, in this case 36}, which means a RET_SMALL stack frame (see includes/rts/storage/ClosureTypes.h for a list of closure types, but make sure you are looking at the right version of this file for the build you're using, because the types do change).

Useful flags

See -ddump-stg, -ddump-simpl, -ddump-cmm, -dppr-debug. For a really close correspondence to the assembly, see -ddump-opt-cmm. You can also use -ddump-to-file to have this output be put into a file rather than to standard output.

Useful hints

Use the dpc macro if you want to avoid the pain of matching the pc address with instructions at every program step:

(gdb) dpc
(gdb) si
0x0000000000408a37 in base_GHCziTopHandler_lvl9_info ()
1: x/i $pc  0x408a37 <base_GHCziTopHandler_lvl9_info+7>:        jb     0x408a55 <base_GHCziTopHandler_lvl9_info+37>
(gdb)
0x0000000000408a39 in base_GHCziTopHandler_lvl9_info ()
1: x/i $pc  0x408a39 <base_GHCziTopHandler_lvl9_info+9>:        mov    $0x5cf248,%ebx
(gdb)
0x0000000000408a3e in base_GHCziTopHandler_lvl9_info ()
1: x/i $pc  0x408a3e <base_GHCziTopHandler_lvl9_info+14>:       movq   $0x408a80,0xfffffffffffffff8(%rbp)
(gdb)

Also note that an empty gdb request repeats the previous command, in this case si.

Sometimes you need to look and step through the compiled code of Haskell libraries. (ToDo: the following is out of date, and refers to the old build system. We should now use -ddump-to-file.) I find it useful to change the lines 43 and 48 of file mk/suffix.mk as follows:
```
43:  $(HC) $(HC_OPTS) -c $< -o $@  -ohi $(basename $@).$(way_)hi -ddump-simpl -ddump-cmm -ddump-stg > $@.output
48:  $(HC) $(HC_OPTS) -c $< -o $@  -ohi $(basename $@).$(way_)hi -ddump-simpl -ddump-cmm -ddump-stg > $@.output
```
Now, I can use file libraries/base/GHC/TopHandler.o.output to make sense of what is going on in TopHandler.o.

Mapping back to the STG code

To figure out which bit of Haskell code corresponds to the assembly fragment you're looking at, you need to look at the STG intermediate code generated by GHC. Use the -ddump-stg flag. The reason we have to look at the STG is because this is the last phase before code generation, after all the transformations have happened, and the symbol names in STG correspond pretty directly to the symbols you see in the object code.

For example, if you're stopped in a code fragment s28a_info, then hopefully you can use info line to see what s28a_info is. You should also be able to search for s28a in the STG output and find it - but you first have to find which module it comes from, and the best way to do that is to grep for s28a in all the modules of your program.

Setting up gdbinit

As GHC's allocation of registers will vary depending on what architecture you're building for, you'll need to modify some values in your gdbinit to make things work for you. In particular, you need to know what values BaseReg and Sp are. You can find them out by looking at includes/stg/MachRegs.h (as well as all of the other registers.) In particular, on x86-32, BaseReg is in ebx and Sp is in ebp, while on x84-64, BaseReg is in r13 and Sp is in rbp.

When you've found & fixed it

Tell someone, they'll probably be impressed :-)
Try to think about whether adding an assertion somewhere might have helped you find the bug quicker, and if so, add one. Test it. Add some commentary next to the assertion to help someone else understand the problem later.
Add a test to the testsuite for your bug.