gccgo fails to compile tomb.go on arm64

[Michael Hudson-Doyle]

Hi, on am2:

ubuntu@arm64:~/gopath/src/launchpad.net/tomb$ gccgo -g -c -o tomb.o tomb.go
tomb.go: In function 'tomb.Wait.pN12_go.tomb.Tomb':
tomb.go:108:1: error: total size of local objects too large
 func (t *Tomb) Wait() error {
 ^
tomb.go: In function 'tomb.Kill.pN12_go.tomb.Tomb':
tomb.go:134:1: error: total size of local objects too large
 func (t *Tomb) Kill(reason error) {
 ^

This is the first error line: http://bazaar.launchpad.net/~niemeyer/tomb/trunk/view/head:/tomb.go#L108 which seems pretty innocuous.

Will dig (and probably kick upstream) in January.

[Michael Hudson-Doyle (mwhudson)]

This is the traceback:

(gdb) bt
#0 0x0000000000b6cf58 in error_at(unsigned int, char const*, ...) ()
#1 0x00000000006d995c in frame_offset_overflow(long, tree_node*) ()
#2 0x00000000006d9b90 in assign_stack_local_1(machine_mode, long, int, int) ()
#3 0x00000000006dcc98 in assign_parms(tree_node*) ()
#4 0x00000000006de6e8 in expand_function_start(tree_node*) ()
#5 0x0000000000600048 in gimple_expand_cfg() ()
#6 0x00000000007b1edc in execute_one_pass(opt_pass*) ()
#7 0x00000000007b225c in execute_pass_list(opt_pass*) ()
#8 0x000000000061b498 in expand_function(cgraph_node*) ()
#9 0x000000000061cbc8 in compile() ()
#10 0x000000000061ceac in finalize_compilation_unit() ()
#11 0x000000000056093c in Gogo::write_globals() ()
#12 0x000000000083f110 in compile_file() ()
#13 0x0000000000840c6c in toplev_main(int, char**) ()

(looks like I need to build a gcc with debugging info though?)

[Michael Hudson-Doyle (mwhudson)]

Hi,

This slightly strangely written program (it's distilled down from frame_offset_overflow in the gcc source itself) should print "bigger" if the first argument is bigger than 10 (or negative, but let's ignore that please):

#include <stdlib.h>
#include <stdio.h>

int a[2] = { 10, 20 };

int
is_bigger (long offset, int index)
{
  unsigned long size = -offset;

  if (size > a[index])
    {
      printf("bigger\n");
      return 1;
    }

  return 0;
}

int
main (int argc, char** argv)
{
  long v;
  v = atol(argv[1]);
  is_bigger(-v, 0);
  return 0;
}

When compiled at -O1 or above (and with inlining disabled at -O2 and above), though, it bungles the 0 case:

(t-doko)mwhudson@arm64:~$ gcc-4.9 -O3 test.c -o test -fno-inline -Wall
(t-doko)mwhudson@arm64:~$ ./test 1
(t-doko)mwhudson@arm64:~$ ./test 11
bigger
(t-doko)mwhudson@arm64:~$ ./test 0
bigger
(t-doko)mwhudson@arm64:~$ gcc-4.9 -O0 test.c -o test -Wall
(t-doko)mwhudson@arm64:~$ ./test 1
(t-doko)mwhudson@arm64:~$ ./test 11
bigger
(t-doko)mwhudson@arm64:~$ ./test 0
(t-doko)mwhudson@arm64:~$

What's going on? Here's the disassembly of is_bigger (at O3):

0000000000400608 <is_bigger>:
  400608: b0000082 adrp x2, 411000 <_GLOBAL_OFFSET_TABLE_+0x28>
  40060c: 91010042 add x2, x2, #0x40
  400610: a9bf7bfd stp x29, x30, [sp,#-16]!
  400614: 52800003 mov w3, #0x0 // #0
  400618: 910003fd mov x29, sp
  40061c: b8a1d841 ldrsw x1, [x2,w1,sxtw #2]
  400620: ab00003f cmn x1, x0
  400624: 540000a2 b.cs 400638 <is_bigger+0x30>
  400628: 90000000 adrp x0, 400000 <_init-0x3f8>
  40062c: 911b6000 add x0, x0, #0x6d8
  400630: 97ffff90 bl 400470 <puts@plt>
  400634: 52800023 mov w3, #0x1 // #1
  400638: 2a0303e0 mov w0, w3
  40063c: a8c17bfd ldp x29, x30, [sp],#16
  400640: d65f03c0 ret

Basically it seems that the condition "-offset > val" is being compiled as "val + offset does not overflow", which is not valid for offset == 0.

[Michael Hudson-Doyle (mwhudson)]

After some considerable prodding and poking, this seems to be a gcc bug: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59744

[Pinskia (pinskia)]

(insn 14 13 15 2 (set (reg:CC_SWP 66 cc)
        (compare:CC_SWP (neg:DI (reg:DI 0 x0 [ offset ]))
            (reg:DI 1 x1 [orig:85 D.3895 ] [85]))) t7.c:11 114 {*compare_negdi}
     (expr_list:REG_DEAD (reg:DI 1 x1 [orig:85 D.3895 ] [85])
        (expr_list:REG_DEAD (reg:DI 0 x0 [ offset ])
            (nil))))

--- CUT ---
Here is a testcase that fails at -O1 and above without any arguments.

int a[2] = { 10, 20 };

int
is_bigger (long, int) __attribute__((noinline,noclone));

int
is_bigger (long offset, int index)
{
  unsigned long size = -offset;

  if (size > a[index])
   return 1;

  return 0;
}

int
main (int argc, char** argv)
{
  long v;
  if (is_bigger(0, 0))
    __builtin_abort ();
  if (!is_bigger(1, 0))
    __builtin_abort ();
  if (is_bigger(-10, 0))
    __builtin_abort ();
  if (!is_bigger(10, 0))
    __builtin_abort ();
  return 0;
}

[Rearnsha (rearnsha)]

Author: rearnsha
Date: Fri Jan 10 15:21:21 2014
New Revision: 206530

URL: http://gcc.gnu.org/viewcvs?rev=206530&root=gcc&view=rev
Log:
PR target/59744
Fix ChangeLog typos in previous commit (r206529).

Modified:
    trunk/gcc/ChangeLog

[Rearnsha (rearnsha)]

Fixed on trunk

[Michael Hudson-Doyle (mwhudson)]

Hi, thanks for the super fast fix. Could it be backported to 4.8? git cherry-pick gives a conflict in aarch64.md which is probably easy to fix if you know how this code works:

(define_insn "*compare_neg<mode>"
<<<<<<< HEAD
  [(set (reg:CC CC_REGNUM)
 (compare:CC
  (match_operand:GPI 0 "register_operand" "r")
  (neg:GPI (match_operand:GPI 1 "register_operand" "r"))))]
||||||| parent of 46b590a... PR target/9744
  [(set (reg:CC_SWP CC_REGNUM)
 (compare:CC_SWP
  (neg:GPI (match_operand:GPI 0 "register_operand" "r"))
  (match_operand:GPI 1 "register_operand" "r")))]
=======
  [(set (reg:CC_Z CC_REGNUM)
 (compare:CC_Z
  (neg:GPI (match_operand:GPI 0 "register_operand" "r"))
  (match_operand:GPI 1 "register_operand" "r")))]
>>>>>>> 46b590a... PR target/9744
  ""
  "cmn\\t%<w>0, %<w>1"
  [(set_attr "v8type" "alus")
   (set_attr "mode" "<MODE>")]
)

[Rearnsha (rearnsha)]

I don't expect that pattern to match on 4.8 (the pattern is not canonical form), which is why this wasn't seen before.

If you can find a test-case that triggers on that branch, I'll do a back-port; otherwise, there's no point.

[Ktkachov (ktkachov)]

I couldn't reproduce the failure using 4.8.

[Michael Hudson-Doyle (mwhudson)]

I saw the problem with Linaro GCC 4.8, but haven't tried the vanilla 4.8 branch. If the committed test case doesn't fail, I'll believe that it is not a problem. Sorry for the noise.

[Matthias Klose (doko)]

fixed in trusty

[Yroux (yroux)]

Author: yroux
Date: Mon Aug 11 22:08:03 2014
New Revision: 213842

URL: https://gcc.gnu.org/viewcvs?rev=213842&root=gcc&view=rev
Log:
gcc/
2014-08-11 Michael Collison <email address hidden>

 Backport from trunk r206529, r206530
 2014-01-10 Richard Earnshaw <email address hidden>

 PR target/59744
 * aarch64-modes.def (CC_Zmode): New flags mode.
 * aarch64.c (aarch64_select_cc_mode): Only allow NEG when the condition
 represents an equality.
 (aarch64_get_condition_code): Handle CC_Zmode.
 * aarch64.md (compare_neg<mode>): Restrict to equality operations.

gcc/testsuite/
2014-08-11 Michael Collison <email address hidden>

 Backport from trunk r206529
 2014-01-10 Richard Earnshaw <email address hidden>

 PR target/59744
 * gcc.target/aarch64/cmn-neg.c: Use equality comparisons.
 * gcc.target/aarch64/cmn-neg2.c: New test.

Added:
    branches/linaro/gcc-4_8-branch/gcc/testsuite/gcc.target/aarch64/cmn-neg2.c
Modified:
    branches/linaro/gcc-4_8-branch/gcc/ChangeLog.linaro
    branches/linaro/gcc-4_8-branch/gcc/config/aarch64/aarch64-modes.def
    branches/linaro/gcc-4_8-branch/gcc/config/aarch64/aarch64.c
    branches/linaro/gcc-4_8-branch/gcc/config/aarch64/aarch64.md
    branches/linaro/gcc-4_8-branch/gcc/testsuite/ChangeLog.linaro
    branches/linaro/gcc-4_8-branch/gcc/testsuite/gcc.target/aarch64/cmn-neg.c