9. Debugging compiled programs

Since the 7.10 release GHC can emit debugging information to help debugging tools understand the code that GHC produces. This debugging information is usable by most UNIX debugging tools.

-g
-g⟨n⟩
Since:7.10, numeric levels since 8.0
Implies:-fexpose-internal-symbols when ⟨n⟩ >= 2.

Emit debug information in object code. Currently only DWARF debug information is supported on x86-64 and i386. Currently debug levels 0 through 3 are accepted:

  • -g0: no debug information produced
  • -g1: produces stack unwinding records for top-level functions (sufficient for basic backtraces)
  • -g2: produces stack unwinding records for top-level functions as well as inner blocks (allowing more precise backtraces than with -g1).
  • -g3: produces GHC-specific DWARF information for use by more sophisticated Haskell-aware debugging tools (see Debugging information entities for details)

If ⟨n⟩ is omitted, level 2 is assumed.

Note that for stack unwinding to be reliable, all libraries, including foreign libraries and those shipped with GHC such as base, must be compiled with unwinding information. GHC binary distributions configured in this way are provided for a select number of platforms; other platforms are advised to build using Hadrian’s +debug_info flavour transformer. Note as well that the built-in unwinding support provided by the base library’s GHC.ExecutionStack module requires that the runtime system be built with libdw support enabled (using the --enable-dwarf-unwind flag to configure while building the compiler) and a platform which libdw supports.

9.1. Tutorial

Let’s consider a simple example,

1
2
3
4
5
6
7
8
 -- fib.hs
 fib :: Int -> Int
 fib 0 = 0
 fib 1 = 1
 fib n = fib (n-1) + fib (n-2)

 main :: IO ()
 main = print $ fib 50

Let’s first see how execution flows through this program. We start by telling GHC that we want debug information,

$ ghc -g -rtsopts fib.hs

Here we used the -g option to inform GHC that it should add debugging information in the produced binary. There are three levels of debugging output: -g0 (no debugging information, the default), -g1 (sufficient for basic backtraces), -g2 (or just -g for short; emitting everything GHC knows). Note that this debugging information does not affect the optimizations performed by GHC.

Tip

Under Mac OS X debug information is kept apart from the executable. After compiling the executable you’ll need to use the dsymutil utility to extract the debugging information and place them in the debug archive,

$ dsymutil fib

This should produce a file named fib.dSYM.

Now let’s have a look at the flow of control. For this we can just start our program under gdb (or an equivalent debugger) as we would any other native executable,

$ gdb --args ./Fib +RTS -V0
Reading symbols from Fib...done.
(gdb) run
Starting program: /opt/exp/ghc/ghc-dwarf/Fib
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
^C
Program received signal SIGINT, Interrupt.
0x000000000064fc7c in cfy4_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:424
424     minusInteger x y = inline plusInteger x (inline negateInteger y)
(gdb)

Here we have used the runtime system’s -V0 option to disable the RTS’s periodic timer which may interfere with our debugging session. Upon breaking into the program gdb shows us a location in our source program corresponding to the current point of execution.

Moreover, we can ask gdb to tell us the flow of execution that lead us to this point in the program,

(gdb) bt
#0  0x000000000064fc7c in cfy4_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:424
#1  0x00000000006eb0c0 in ?? ()
#2  0x000000000064301c in cbuV_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:323
#3  0x000000000064311b in integerzmgmp_GHCziIntegerziType_eqInteger_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:312
#4  0x0000000000406eca in roz_info () at Fib.hs:2
#5  0x00000000006eb0c0 in ?? ()
#6  0x000000000064f075 in cfru_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:412
#7  0x00000000006eb0c0 in ?? ()
#8  0x000000000064f075 in cfru_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:412
#9  0x00000000006eb0c0 in ?? ()
#10 0x000000000064eefe in integerzmgmp_GHCziIntegerziType_plusInteger_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:393
...
#64 0x0000000000643ac8 in integerzmgmp_GHCziIntegerziType_ltIntegerzh_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:343
#65 0x00000000004effcc in base_GHCziShow_zdwintegerToString_info () at libraries/base/GHC/Show.hs:443
#66 0x00000000004f0795 in base_GHCziShow_zdfShowIntegerzuzdcshow_info () at libraries/base/GHC/Show.hs:145
#67 0x000000000048892b in cdGW_info () at libraries/base/GHC/IO/Handle/Text.hs:595
#68 0x0000000000419cb2 in base_GHCziBase_thenIO1_info () at libraries/base/GHC/Base.hs:1072

Hint

Here we notice the first bit of the stack trace has many unidentified stack frames at address 0x006eb0c0. If we ask gdb about this location, we find that these frames are actually STG update closures,

(gdb) print/a 0x006eb0c0
$1 = 0x6eb0c0 <stg_upd_frame_info>

The reason gdb doesn’t show this symbol name in the backtrace output is an infidelity in its interpretation of debug information, which assumes an invariant preserved in C but not Haskell programs. Unfortunately it is necessary to work around this manually until this behavior is fixed upstream.

Note

Because of the aggressive optimization that GHC performs to the programs it compiles it is quite difficult to pin-point exactly which point in the source program a given machine instruction should be attributed to. In fact, internally GHC associates each instruction with a set of source locations. When emitting the standard debug information used by gdb and other language-agnostic debugging tools, GHC is forced to heuristically choose one location from among this set.

For this reason we should be cautious when interpreting the source locations provided by GDB. While these locations will usually be in some sense “correct”, they aren’t always useful. This is why profiling tools targeting Haskell should supplement the standard source location information with GHC-specific annotations (emitted with -g2) when assigning costs.

Indeed, we can even set breakpoints,

(gdb) break fib.hs:4
Breakpoint 1 at 0x406c60: fib.hs:4. (5 locations)
(gdb) run
Starting program: /opt/exp/ghc/ghc-dwarf/Fib

Breakpoint 1, c1RV_info () at Fib.hs:4
4        fib n = fib (n-1) + fib (n-2)
(gdb) bt
#0  c1RV_info () at Fib.hs:4
#1  0x00000000006eb0c0 in ?? ()
#2  0x0000000000643ac8 in integerzmgmp_GHCziIntegerziType_ltIntegerzh_info () at libraries/integer-gmp/src/GHC/Integer/Type.hs:343
#3  0x00000000004effcc in base_GHCziShow_zdwintegerToString_info () at libraries/base/GHC/Show.hs:443
#4  0x00000000004f0795 in base_GHCziShow_zdfShowIntegerzuzdcshow_info () at libraries/base/GHC/Show.hs:145
#5  0x000000000048892b in cdGW_info () at libraries/base/GHC/IO/Handle/Text.hs:595
#6  0x0000000000419cb2 in base_GHCziBase_thenIO1_info () at libraries/base/GHC/Base.hs:1072
#7  0x00000000006ebcb0 in ?? () at rts/Exception.cmm:332
#8  0x00000000006e7320 in ?? ()
(gdb)

Due to the nature of GHC’s heap and the heavy optimization that it performs, it is quite difficult to probe the values of bindings at runtime. In this way, the debugging experience of a Haskell program with DWARF support is still a bit impoverished compared to typical imperative debuggers.

9.2. Requesting a stack trace from Haskell code

GHC’s runtime system has built-in support for collecting stack trace information from a running Haskell program. This currently requires that the libdw library from the elfutils package is available. Of course, the backtrace will be of little use unless debug information is available in the executable and its dependent libraries.

Stack trace functionality is exposed for use by Haskell programs in the GHC.ExecutionStack module. See the Haddock documentation in this module for details regarding usage.

9.3. Requesting a stack trace with SIGQUIT

On POSIX-compatible platforms GHC’s runtime system (when built with libdw support) will produce a stack trace on stderr when a SIGQUIT signal is received (on many systems this signal can be sent using Ctrl-\). For instance (using the same fib.hs as above),

$ ./fib  &  killall -SIGQUIT fib

Caught SIGQUIT; Backtrace:
0x7f3176b15dd8    dwfl_thread_getframes (/usr/lib/x86_64-linux-gnu/libdw-0.163.so)
0x7f3176b1582f    (null) (/usr/lib/x86_64-linux-gnu/libdw-0.163.so)
0x7f3176b15b57    dwfl_getthreads (/usr/lib/x86_64-linux-gnu/libdw-0.163.so)
0x7f3176b16150    dwfl_getthread_frames (/usr/lib/x86_64-linux-gnu/libdw-0.163.so)
      0x6dc857    libdwGetBacktrace (rts/Libdw.c:248.0)
      0x6e6126    backtrace_handler (rts/posix/Signals.c:541.0)
0x7f317677017f    (null) (/lib/x86_64-linux-gnu/libc-2.19.so)
      0x642e1c    integerzmgmp_GHCziIntegerziType_eqIntegerzh_info (libraries/integer-gmp/src/GHC/Integer/Type.hs:320.1)
      0x643023    integerzmgmp_GHCziIntegerziType_eqInteger_info (libraries/integer-gmp/src/GHC/Integer/Type.hs:312.1)
      0x406eca    roz_info (/opt/exp/ghc/ghc-dwarf//Fib.hs:2.1)
      0x6eafc0    stg_upd_frame_info (rts/Updates.cmm:31.1)
      0x64ee06    integerzmgmp_GHCziIntegerziType_plusInteger_info (libraries/integer-gmp/src/GHC/Integer/Type.hs:393.1)
      0x6eafc0    stg_upd_frame_info (rts/Updates.cmm:31.1)
...
      0x6439d0    integerzmgmp_GHCziIntegerziType_ltIntegerzh_info (libraries/integer-gmp/src/GHC/Integer/Type.hs:343.1)
      0x4efed4    base_GHCziShow_zdwintegerToString_info (libraries/base/GHC/Show.hs:442.1)
      0x4f069d    base_GHCziShow_zdfShowIntegerzuzdcshow_info (libraries/base/GHC/Show.hs:145.5)
      0x488833    base_GHCziIOziHandleziText_zdwa8_info (libraries/base/GHC/IO/Handle/Text.hs:582.1)
      0x6ebbb0    stg_catch_frame_info (rts/Exception.cmm:370.1)
      0x6e7220    stg_stop_thread_info (rts/StgStartup.cmm:42.1)

9.4. Implementor’s notes: DWARF annotations

Note

Most users don’t need to worry about the details described in this section. This discussion is primarily targeted at tooling authors who need to interpret the GHC-specific DWARF annotations contained in compiled binaries.

When invoked with the -g flag GHC will produce standard DWARF v4 debugging information. This format is used by nearly all POSIX-compliant targets and can be used by debugging and performance tools (e.g. gdb, lldb, and perf) to understand the structure of GHC-compiled programs.

In particular GHC produces the following DWARF sections,

.debug_info
Debug information entities (DIEs) describing all of the basic blocks in the compiled program.
.debug_line

Line number information necessary to map instruction addresses to line numbers in the source program.

Note that the line information in this section is not nearly as rich as the information provided in .debug_info. Whereas .debug_line requires that each instruction is assigned exactly one source location, the DIEs in .debug_info can be used to identify all relevant sources locations.

.debug_frames
Call frame information (CFI) necessary for stack unwinding to produce a call stack trace.
.debug_arange
Address range information necessary for efficient lookup in debug information.

9.4.1. Debugging information entities

GHC may produce the following standard DIEs in the .debug_info section,

DW_TAG_compile_unit
Represents a compilation unit (e.g. a Haskell module).
DW_TAG_subprogram
Represents a C-- top-level basic block.
DW_TAG_lexical_block
Represents a C-- basic block. Note that this is a slight departure from the intended meaning of this DIE type as it does not necessarily reflect lexical scope in the source program.

As GHC’s compilation products don’t map perfectly onto DWARF constructs, GHC takes advantage of the extensibility of the DWARF standard to provide additional information.

Unfortunately DWARF isn’t expressive enough to fully describe the code that GHC produces. This is most apparent in the case of line information, where GHC is forced to choose some between a variety of possible originating source locations. This limits the usefulness of DWARF information with traditional statistical profiling tools. For profiling it is recommended that one use the extended debugging information. See the Profiling section below.

In addition to the usual DIEs specified by the DWARF specification, GHC produces a variety of others using the vendor-extensibility regions of the tag and attribute space.

9.4.1.1. DW_TAG_ghc_src_note

DW_TAG_ghc_src_note DIEs (tag 0x5b01) are found as children of DW_TAG_lexical_block DIEs. They describe source spans which gave rise to the block; formally these spans are causally responsible for produced code: changes to code in the given span may change the code within the block; conversely changes outside the span are guaranteed not to affect the code in the block.

Spans are described with the following attributes,

DW_AT_ghc_span_file (0x2b00, string)
the name of the source file
DW_AT_ghc_span_start_line (0x2b01, integer)
the line number of the beginning of the span
DW_AT_ghc_span_start_col (0x2b02, integer)
the column number of the beginning of the span
DW_AT_ghc_span_end_line (0x2b03, integer)
the line number of the end of the span
DW_AT_ghc_span_end_col (0x2b04, integer)
the column number of the end of the span

9.5. Further Reading

For more information about the debug information produced by GHC see Peter Wortmann’s PhD thesis, *Profiling Optimized Haskell: Causal Analysis and Implementation*.

9.6. Direct Mapping

In addition to the DWARF debug information, which can be used by many standard tools, there is also a GHC specific way to map info table pointers to a source location. This lookup table is generated by using the -finfo-table-map flag.

-finfo-table-map
Since:9.2

This flag enables the generation of a table which maps the address of an info table to an approximate source position of where that info table statically originated from. If you also want more precise information about constructor info tables then you should also use -fdistinct-constructor-tables.

The -finfo-table-map flag will increase the binary size by quite a lot, depending on how big your project is. For compiling a project the size of GHC the overhead was about 200 megabytes.

Since:9.8

If you wish to reduce the size of -finfo-table-map enabled binaries, consider building GHC from source and supplying the --enable-ipe-data-compression flag to the configure script. This will cause GHC to compress the -finfo-table-map related debugging information included in binaries using the libzstd compression library. Note: This feature requires that the machine building GHC has libzstd installed. The compression library libzstd may optionally be statically linked in the resulting compiler (on non-darwin machines) using the --enable-static-libzstd configure flag.

In a test compiling GHC itself, the size of the -finfo-table-map enabled build results was reduced by over 20% when compression was enabled.

Since:9.10
Implies:-finfo-table-map-with-stack
Implies:-finfo-table-map-with-fallback
-finfo-table-map-with-stack
Since:9.10

Include info tables for STACK closures in the info table map. Note that this flag is implied by -finfo-table-map.

-fno-info-table-map-with-stack
Since:9.10

STACK info tables are often the majority of entries in the info table map. However, despite their contribution to the executable size, they are rarely useful unless debugging with a tool such as ghc-debug. Use this flag to omit STACK info tables from the info table map and decrease the size of executables with info table profiling information.

-finfo-table-map-with-fallback
Since:9.10

Include info tables with no source location information in the info table map. Note that this flag is implied by -finfo-table-map.

-fno-info-table-map-with-fallback
Since:9.10

Some info tables, such as those for primitive closure types, will have no provenance location in the program source. With -finfo-table-map, those info tables are given default source locations and included in the info table map. Use this flag to omit them from the info table map and decrease the size of executables with info table profiling information.

-fdistinct-constructor-tables
Since:9.2

For every usage of a data constructor in the source program a new info table will be created. This is useful with -finfo-table-map and the -hi profiling mode as each info table will correspond to the usage of a data constructor rather than the data constructor itself.

9.7. Querying the Info Table Map

If it is generated then the info table map can be used in two ways.

  1. The whereFrom Haskell function can be used to determine the source position which we think a specific closure was created.
  2. The complete mapping is also dumped into the eventlog.

If you are using gdb then you can use the lookupIPE function (provided by IPE.h and exported in the public API) directly in order to find any information which is known about the info table for a specific closure.