RE: Miscompilation...

["Dave Korn" <dave dot korn at artimi dot com>]

On 09 April 2007 02:51, Bill Wendling wrote:

> (on my iMac). The problem seems to be that, in the TConvertType
> function passes
> "type->precision" as an HI instead of SI to the "bork" function. The
> asm code is:
> 
>          movl $1104150528, (%esp)
>          movl    $0, -4(%ebp)
>          movl    8(%ebp), %eax
>          movzwl  (%eax), %eax
>          andw    $511, %ax
>          movw    %ax, 4(%esp)
>          movl    -4(%ebp), %eax
>          movl    %eax, (%esp)
>          call    __Z4borkPKvj
> 
> for TConvertType, so the %ax isn't putting a full SI onto the stack,
> leaving
> garbage in the MSBs of 4(%esp), which "bork" expects to be 0, of course.

  Confirmed on x86/cygwin.  I removed the asms just in case they were somehow
confusing the compiler but the generated code is still clearly incorrect.

004010a2 <__Z12TConvertTypeP9tree_type>:
  4010a2:       55                      push   %ebp
  4010a3:       89 e5                   mov    %esp,%ebp
  4010a5:       83 ec 18                sub    $0x18,%esp
  4010a8:       c7 45 fc 00 00 00 00    movl   $0x0,0xfffffffc(%ebp)
  4010af:       8b 45 08                mov    0x8(%ebp),%eax
  4010b2:       0f b7 00                movzwl (%eax),%eax
  4010b5:       66 25 ff 01             and    $0x1ff,%ax
  4010b9:       66 89 44 24 04          mov    %ax,0x4(%esp)
  4010be:       8b 45 fc                mov    0xfffffffc(%ebp),%eax
  4010c1:       89 04 24                mov    %eax,(%esp)
  4010c4:       e8 87 ff ff ff          call   401050 <__Z4borkPKvj>
  4010c9:       c9                      leave
  4010ca:       c3                      ret
  4010cb:       90                      nop

  Looking at tree dumps for the two functions ConvertType and foo,

;; Function const void* foo(tree_type*) (_Z3fooP9tree_type)

const void* foo(tree_type*) (type)
{
  unsigned int S;
  const void * Ty;
  void * D.2934;
  const void * D.2933;
  <unnamed type> D.2932;

  # BLOCK 2
  # PRED: ENTRY (fallthru)
  Ty = 0B;
  D.2932 = type->precision;
  S = (unsigned int) D.2932;
  D.2934 = bork (Ty, S);
  D.2933 = D.2934;
  return D.2933;
  # SUCC: EXIT

}

;; Function const void* TConvertType(tree_type*) (_Z12TConvertTypeP9tree_type)

const void* TConvertType(tree_type*) (type)
{
  const void * Ty;
  void * D.2926;
  <unnamed type> D.2925;
  const void * D.2924;

  # BLOCK 2
  # PRED: ENTRY (fallthru)
  Ty = 0B;
  D.2925 = type->precision;
  D.2926 = bork (Ty, D.2925);
  D.2924 = D.2926;
  return D.2924;
  # SUCC: EXIT

}

the difference is just in whether the <unnamed type> representing the bitfield
is passed directly to bork as a parameter; when forcibly cast via an unsigned,
it gets it right.  The generated RTL appears to then select a minimal size for
<unsigned type> and not take account of argument promotion:


;; Function const void* TConvertType(tree_type*) (_Z12TConvertTypeP9tree_type)


;; Generating RTL for tree basic block 2

;; Ty = 0B
(insn 7 5 0 (set (mem/f/c/i:SI (plus:SI (reg/f:SI 54 virtual-stack-vars)
                (const_int -4 [0xfffffffc])) [0 Ty+0 S4 A32])
        (const_int 0 [0x0])) -1 (nil)
    (nil))

;; D.2925 = type->precision
(insn 9 7 11 (set (reg/f:SI 62)
        (mem/f/c/i:SI (reg/f:SI 53 virtual-incoming-args) [0 type+0 S4 A32]))
-1 (nil)
    (nil))

(insn 11 9 0 (parallel [
            (set (reg:HI 59 [ D.2925 ])
                (and:HI (mem/s:HI (reg/f:SI 62) [0 S2 A32])
                    (const_int 511 [0x1ff])))
            (clobber (reg:CC 17 flags))
        ]) -1 (nil)
    (nil))

;; D.2926 = bork (Ty, D.2925)
(insn 12 11 13 (set (mem:HI (plus:SI (reg/f:SI 56 virtual-outgoing-args)
                (const_int 4 [0x4])) [0 S2 A32])
        (reg:HI 59 [ D.2925 ])) -1 (nil)
    (nil))


> I believe I got the TOT -- .svn/entries says "svn://gcc.gnu.org/svn/
> gcc/trunk". Is this a known problem?

  It's the first I've heard of it, please file a PR.  I'm slightly amazed this
hasn't been causing bootstrap breakage already, it looks really wrong.

    cheers,
      DaveK
-- 
Can't think of a witty .sigline today....