GCC Bugzilla Bug 25196

I really had a hard time tracking this down from a large program: gcc 4.0.2 on
i386-linux sometimes passes wrong arguments.

Please try the attached source file on i386 with "gcc-4.0.2 -march=i386 -O3
-fomit-frame-pointer".

When looking at the disassembly of f2() it seems that %edi is incorrectly used
twice:

  5e:   89 f8                   mov    %edi,%eax   !!!!!
  60:   ff 74 24 24             pushl  0x24(%esp)
  64:   55                      push   %ebp
  65:   89 f9                   mov    %edi,%ecx   !!!!!
  67:   89 da                   mov    %ebx,%edx
  69:   e8 92 ff ff ff          call   0 <f1>

This wrong-code bug is probably a 3.2, 3.3 and 3.4 regression.

~Markus

Created an attachment (id=10373) [edit]
wrong_args.c

Please try "gcc-4.0.2 -march=i386 -O3 -fomit-frame-pointer wrong_args.c"

This works at least on the mainline (but I don't know if this is a latent bug
or not).

Serious wrong code problem: P1.

// short testcase, compile with "-m32 -march=i386 -O3 -fomit-frame-pointer"

extern void abort (void);

static int j;

static void __attribute__((noinline))
f1 (int a, int b, int c, int d, int e)
{
  j = a;
}

int __attribute__((noinline))
f2 (int a, int b, int c, int d, int e)
{
  if ((b & 0x1111) != 1)
    f1 (a, b, c, d, e);
  return 0;
}

int main (void)
{
  f2 (123, 0, 0, 0, 0);
  if (j != 123)
    abort ();
  return 0;
}

---

Note that this bug disappears with -fno-gcse-after-reload, here is difference
in generated assembly between -fno-gcse-after-reload (---) and
-fgcse-after-reload (+++):

$ diff -U 5 1.s 2.s                                                             
--- 1.s 2005-12-20 12:10:44.000000000 +0300
+++ 2.s 2005-12-20 12:10:41.000000000 +0300
@@ -14,13 +14,13 @@
        movl    12(%esp), %ecx
        movl    %edx, %eax
        andl    $4369, %eax
        decl    %eax
        je      .L4
+       movl    %ecx, %eax
        pushl   20(%esp)
        pushl   20(%esp)
-       movl    12(%esp), %eax
        call    f1
        popl    %eax
        popl    %edx
 .L4:
        xorl    %eax, %eax
$ 

So this bug looks very much like bug 23453.

Re. comment #4: but this new PR has a much simpler test case :-)

*** Bug 23453 has been marked as a duplicate of this bug. ***

Does not fail with trunk or the head of the gcc 4.1 branch.  But it does fail
with gcc 4.0.2.  I'm going to try it with the head of the gcc 4.0 branch now.

Gross.  According to a comment in postreload.c:move2add_note_store(), we can
have pushes without REG_INC notes:
  /* Some targets do argument pushes without adding REG_INC notes.  */

So we need to go look for those {pre,post}{decrements,increments} ourselves. 
With this patch, we just do what postreload.c does.


Index: postreload-gcse.c
===================================================================
--- postreload-gcse.c   (revision 108853)
+++ postreload-gcse.c   (working copy)
@@ -676,6 +676,17 @@ record_last_set_info (rtx dest, rtx sett
           /* Ignore pushes, they clobber nothing.  */
           && ! push_operand (dest, GET_MODE (dest)))
     record_last_mem_set_info (last_set_insn);
+
+  /* ??? Some targets do argument pushes without adding REG_INC notes.
+     See e.g. PR25196, where a pushsi2 on i386 doesn't have REG_INC
+     notes.  */
+  if (MEM_P (dest))
+    {
+      dest = XEXP (dest, 0);
+      if (GET_CODE (dest) == PRE_INC || GET_CODE (dest) == POST_INC
+         || GET_CODE (dest) == PRE_DEC || GET_CODE (dest) == POST_DEC)
+       record_last_reg_set_info (last_set_insn, REGNO (XEXP (dest, 0)));
+    }
 }

Fixed on the trunk and on the GCC 4.1 branch.
See http://gcc.gnu.org/ml/gcc-cvs/2005-12/msg01177.html and
http://gcc.gnu.org/ml/gcc-cvs/2005-12/msg01178.html (I used the wrong bug
number in the commit >:-/)

Will ask for permission for the 4.0 branch in a few days.

What is the status of this bug for gcc 4.0.3 ?

According to the thread at
http://gcc.gnu.org/ml/gcc-patches/2006-01/msg00542.html it seems it has been
approved a while ago.

Steven, would you please apply this patch ASAP?

Thanks,

-- Mark

Subject: Bug 25196

Author: steven
Date: Mon Feb 27 21:12:54 2006
New Revision: 111490

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=111490
Log:
        Backport from mainline and the GCC 4.1 branch:
        PR rtl-optimization/25196
        * postreload-gcse.c (record_last_set_info): Notice stack pointer
        changes in push insns without REG_INC notes.

Added:
    branches/gcc-4_0-branch/gcc/testsuite/gcc.dg/pr25196.c
Modified:
    branches/gcc-4_0-branch/gcc/ChangeLog
    branches/gcc-4_0-branch/gcc/postreload-gcse.c

'tis done.