[Bug target/47812] New: [x32] __builtin_strlen is miscompiled at -O2

[Bug target/47812] New: [x32] __builtin_strlen is miscompiled at -O2

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47812

           Summary: [x32] __builtin_strlen is miscompiled at -O2
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hjl.tools@gmail.com
                CC: ubizjak@gmail.com


On x32 branch, I got

[hjl@gnu-6 strcat-1]$ cat tester.c
char one[50] = "ijk";
int
main (void)
{
  return __builtin_strlen (one) != 3;
}
[hjl@gnu-6 strcat-1]$ make tester.s
/usr/gcc-4.6.0-x32/bin/gcc -mx32 -g -O2 -S tester.c -o tester.s
[hjl@gnu-6 strcat-1]$ cat tester.s
    .file    "tester.c"
    .text
.Ltext0:
    .section    .text.startup,"ax",@progbits
    .p2align 4,,15
    .globl    main
    .type    main, @function
main:
.LFB0:
    .file 1 "tester.c"
    .loc 1 4 0
    .cfi_startproc
    .loc 1 5 0
    movl    $one, %edx
.L2:
    movl    (%rdx), %eax
    addq    $4, %rdx
    leal    -16843009(%eax), %ecx
    notl    %eax
    andl    %eax, %ecx
    andl    $-2139062144, %ecx
    je    .L2
    .loc 1 6 0
    movl    $1, %eax
    ret
    .cfi_endproc

This always returns 1 instead of 0.

[Bug target/47812] [x32] __builtin_strlen is miscompiled at -O2

[xinliangli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47812

davidxl <xinliangli at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |xinliangli at gmail dot com

--- Comment #1 from davidxl <xinliangli at gmail dot com> 2011-02-19 07:05:47 UTC ---

There is a bug in the expand phase.   Pseudo reg 59 is not initialized which
leads to an initialization to 0 by the init-reg pass. combine pass then happily
simply the comparison result with 1 which is returned.


The bad case with x32 where reg:SI 64 is initialized instead of 59:

insn 25 24 0 (parallel [
            (set (reg:DI 64)
                (minus:DI (reg:DI 64)
                    (reg:DI 62)))
            (clobber (reg:CC 17 flags))
        ]) pr47812.c:5 -1
     (nil))

;; return D.2771_2;

(insn 29 25 30 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg:SI 59 [ D.2772 ])
            (const_int 3 [0x3]))) pr47812.c:5 -1
     (nil))

The good case x86_64:

(insn 24 23 0 (parallel [
            (set (reg:DI 59 [ D.2772 ])
                (minus:DI (reg:DI 59 [ D.2772 ])
                    (reg:DI 62)))
            (clobber (reg:CC 17 flags))
        ]) pr47812.c:5 -1
     (nil))

;; return D.2771_2;

(insn 26 24 27 (set (reg:CCZ 17 flags)
        (compare:CCZ (reg:DI 59 [ D.2772 ])
            (const_int 3 [0x3]))) pr47812.c:5 -1
     (nil))


David

[Bug target/47812] [x32] __builtin_strlen is miscompiled at -O2

[xinliangli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47812

--- Comment #2 from davidxl <xinliangli at gmail dot com> 2011-02-19 07:44:03 UTC ---
In ix86_expand_strlen:

if (GET_MODE (out) != Pmode)
   out = convert_to_mode (Pmode, out, 1);


Here out is REG:SI 59 -- the result register for strlen. After conversion, new
reg:DI 64 is created. There are two problems here:

1) the result is now 64, but the caller 'expand_builtin_strlen' still uses reg
59
2) the conversion itself generates an uninitialized use.

David

[Bug target/47812] [x32] __builtin_strlen is miscompiled at -O2

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47812

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |DUPLICATE

--- Comment #3 from H.J. Lu <hjl.tools at gmail dot com> 2011-02-19 14:22:45 UTC ---
(In reply to comment #2)
> In ix86_expand_strlen:
> 
> if (GET_MODE (out) != Pmode)
>    out = convert_to_mode (Pmode, out, 1);
> 
> 
> Here out is REG:SI 59 -- the result register for strlen. After conversion, new
> reg:DI 64 is created. There are two problems here:
> 
> 1) the result is now 64, but the caller 'expand_builtin_strlen' still uses reg
> 59
> 2) the conversion itself generates an uninitialized use.
> 

Thanks for the analysis.  This is a dup of PR47364. I have a patch.

*** This bug has been marked as a duplicate of bug 47364 ***