c-reduce nano howto

What is c-reduce? The home page gives very nice explanation:

C-Reduce is a tool that takes a large C, C++, or OpenCL file that has a property of interest
(such as triggering a compiler bug) and automatically produces a much smaller C/C++ file that
has the same property. It is intended for use by people who discover and report bugs in compilers
and other tools that process source code.

Yesterday Amynka shared an interesting guile build failure: bug #610052

The build error is of rare and cryptic sort:

../libguile/guile-snarf -o vm.x vm.c -DHAVE_CONFIG_H -DBUILDING_LIBGUILE=1 -I.. -I.. -I../lib -I../lib -I/usr/lib64/libffi-3.2.1/include  -I/var/tmp/portage/dev-scheme/guile-2.0.14/work/guile-2.0.14  -march=native -mtune=native -Os -pipe -ggdb
...
libtool: compile:  gcc -DHAVE_CONFIG_H -DBUILDING_LIBGUILE=1 -I.. -I.. -I../lib -I../lib -I/usr/lib64/libffi-3.2.1/include -I/var/tmp/portage/dev-scheme/guile-2.0.14/work/guile-2.0.14 -pthread -Wall -Wmissing-prototypes -Wdeclaration-after-statement -Wpointer-arith -Wswitch-enum -fno-strict-aliasing -fwrapv -fvisibility=hidden -march=native -mtune=native -Os -pipe -ggdb -c vm.c  -fPIC -DPIC -o .libs/libguile_2.0_la-vm.o
...
Unable to coalesce ssa_names 48 and 3288 which are marked as MUST COALESCE.
sp_48(ab) and  sp_3288(ab)
vm-engine.c: In function 'vm_debug_engine':
vm.c:666:19: internal compiler error: SSA corruption
 #define VM_NAME   vm_debug_engine
...
Please submit a full bug report,
with preprocessed source if appropriate.
See <https://bugs.gentoo.org/> for instructions.

internal compiler error: SSA corruption says the thing we see is a GCC bug.

User provided plenty information like:

I’ve tried to reproduce it on i7-2700K CPU and did not succeed. -march=native is a flag that picks instructions available on currently running CPU model+GCC version.

To find out what AMD A10 7850K translates to I’ve looked at available -march= options in man gcc.

That is a huge list even for x86_64. I had to pick from an extensive AMD subset (starts at -march=k6). Something like -march=amdfam10 or later (-march=bdver1, -march=bdver2, etc.) looked like a good match. I chose -march=bdver2 mostly at random :)

bdver2
    AMD Family 15h core based CPUs with x86-64 instruction set support.  (This supersets BMI, TBM, F16C, FMA, FMA4, AVX, XOP, LWP, AES, PCL_MUL, CX16, MMX, SSE,
    SSE2, SSE3, SSE4A, SSSE3, SSE4.1, SSE4.2, ABM and 64-bit instruction set extensions.)

Reran build command changing -march=native to -march=bdver2

gcc -DHAVE_CONFIG_H -DBUILDING_LIBGUILE=1 -I.. -I.. -I../lib -I../lib -I/usr/lib64/libffi-3.2.1/include -I/var/tmp/portage/dev-scheme/guile-2.0.14/work/guile-2.0.14 -pthread -Wall -Wmissing-prototypes -Wdeclaration-after-statement -Wpointer-arith -Wswitch-enum -fno-strict-aliasing -fwrapv -fvisibility=hidden -march=bdver2 -mtune=bdver2 -Os -pipe -ggdb -c vm.c  -fPIC -DPIC -o .libs/libguile_2.0_la-vm.o

and got the same GCC crash \o/. Yay! Now we have something to look at.

vm.c relies on local (to guile) and system headers. It’s not clear what triggered the error, so I wanted to get a selfcontained example. GCC bug reporting manual suggests using -save-temps option to bundle everything required into a single file. Single files are easier to deal with.

gcc ...<skip>... -march=bdver2 -mtune=bdver2 -Os -pipe -ggdb -c vm.c  -fPIC -DPIC -o .libs/libguile_2.0_la-vm.o -save-temps

The result is stored in vm.i. We can shrink our compile command down to:

gcc -march=bdver2 -mtune=bdver2 -Os -c vm.i

vm.i file is 1052623 bytes long (like a book of decent size). Who knows what went wrong there.

Googling for internal compiler error: SSA corruption revealed quite a few bugs related to SSA bugs. One of bugs hinted at c-reduce as a tool to shrink original C source files.

I’ve decided to give it a try. c-reduce is very easy to use: we need to construct a wrapper script that has 2 exit codes:

# cat test.sh
LANG=C gcc-4.9.4 -march=bdver2 -O1 -c vm.i -o vm.o >gcc_out.txt 2>&1
grep 'internal compiler error' gcc_out.txt >/dev/null 2>&1

That’s it! Let’s start the process:

$ time creduce ./test.sh vm.i
===< 2043 >===
running 4 interestingness tests in parallel
===< pass_includes :: 0 >===
===< pass_unifdef :: 0 >===
===< pass_comments :: 0 >===
(0.0 %, 1052583 bytes)
===< pass_blank :: 0 >===
(0.8 %, 1044718 bytes)
(5.1 %, 998478 bytes)
===< pass_clang_binsrch :: replace-function-def-with-decl >===
(6.4 %, 985456 bytes)
(8.3 %, 965433 bytes)
(8.7 %, 960939 bytes)
...
...<takes a while>...
...
      ******** /tmp/y/vm.i ********

*a;
b() {
  char *c;
  static d;
  e();
  d = c++;
  f();
  _setjmp();
  goto *a[*c];
}

real    7m53,030s
user    19m0,687s
sys     3m37,214s

The final result is stored in original vm.i file:

*a;
b() {
  char *c;
  static d;
  e();
  d = c++;
  f();
  _setjmp();
  goto *a[*c];
}

Only 88 bytes! Not the best piece of C ever written but still perfectly valid. And way better than original size-wise :)

This code is portable as it does not contain any gcc version-specific intrinsics like __builtin_va_list or anything like that.

Now we more or less know what to look for. GCC had a similar bug where -O1 -ftree-vectorize was enough to trigger internal compiler error (aka ICE).

Our case is even simpler. It’s not CPU model specific anymore:

$ gcc-4.9.4 -c vm.i -O1
...
Unable to coalesce ssa_names 2 and 9 which are marked as MUST COALESCE.
c_2(ab) and  c_9(ab)

So, why wasn’t the original bug triggered on my -march=corei7-avx machine? I don’t know! Perhaps GCC’s intermediate representation slightly depends on target CPU model.

None of gcc-5.4.0 or gcc-6.3.0 crash on that small snippet of code. The bug was likely fixed but not merged to gcc-4.9 branch. gcc-4.9.4 was the final release in 4.9 branch.

Random findings:

Have fun!

Posted on March 5, 2017
comments powered by Disqus