Re: optimization/7232: Incorrect code generate for volatiles

Re: optimization/7232: Incorrect code generate for volatiles

[Sven Lundblad]

The following reply was made to PR optimization/7232; it has been noted by GNATS.

From: Sven Lundblad <svlu@enea.se>
To: Franz.Sirl-kernel@lauterbach.com
Cc: gcc-gnats@gcc.gnu.org,
 karm@enea.se,
 steby@enea.se
Subject: Re: optimization/7232: Incorrect code generate for volatiles
Date: Mon, 8 Jul 2002 23:29:48 +0200

 Hi,
 
 Thanks for the fast response! The generated code is still incorrect. Note the 
 two lwz instructtions that loads from the same address (r11+8). Please see my 
 comments below (and please correct me if I'm wrong).
 
 Regards,
 Sven Lundblad
 
  .file   "test.c"
           .section        ".text"
           .align 2
           .type   force_gcc_ppc_bug,@function
  force_gcc_ppc_bug:
           lis 11,ll@ha     /* load address of ll into r11 high */
           li 0,4660
           la 11,ll@l(11)   /* load address of ll into r11 low */
           lwz 10,8(11)    /* load ll.rtFreeList into r10, should be
                                  ll.rtFreeList->next (the lnk variable) ??? */
           lwz 9,8(11)      /* load ll.rtFreeList into r9 */
           stw 0,0(10)     /* store 0x1234 into ll.rtFreeList->dummy WRONG! */
           mr 3,10          /* put ll.rtFreeList into r3 (but should be the lnk
                                  pointer) */
           lwz 0,4(9)       /* load ll.rtFreeList->next into r0 */
           li 9,0
           stw 0,8(11)     /* store ll.rtFreeList->next into ll.rtFreeList */
           li 0,18
           stw 9,4(10)
           stb 0,42(10)    /* store 0x12 into space outside allocated data! */
           blr
  .Lfe1:
           .size   force_gcc_ppc_bug,.Lfe1-force_gcc_ppc_bug
           .comm   ll,12,4
           .ident  "GCC: (GNU) 3.1.1 20020704 (prerelease)"
 

Re: optimization/7232: Incorrect code generate for volatiles

[Franz Sirl]

The following reply was made to PR optimization/7232; it has been noted by GNATS.

From: Franz Sirl <Franz.Sirl-kernel@lauterbach.com>
To: svlu@enea.se
Cc: gcc-gnats@gcc.gnu.org,karm@enea.se,steby@enea.se
Subject: Re: optimization/7232: Incorrect code generate for volatiles
Date: Tue, 09 Jul 2002 10:45:13 +0200

 At 23:29 08.07.2002, Sven Lundblad wrote:
 >Hi,
 >
 >Thanks for the fast response! The generated code is still incorrect. Note the
 >two lwz instructtions that loads from the same address (r11+8). Please see my
 >comments below (and please correct me if I'm wrong).
 >
 >Regards,
 >Sven Lundblad
 >
 >  .file   "test.c"
 >           .section        ".text"
 >           .align 2
 >           .type   force_gcc_ppc_bug,@function
 >  force_gcc_ppc_bug:
 >           lis 11,ll@ha     /* load address of ll into r11 high */
 >           li 0,4660
 >           la 11,ll@l(11)   /* load address of ll into r11 low */
 >           lwz 10,8(11)    /* load ll.rtFreeList into r10, should be
 >                                  ll.rtFreeList->next (the lnk variable) 
 > ??? */
 >           lwz 9,8(11)      /* load ll.rtFreeList into r9 */
 >           stw 0,0(10)     /* store 0x1234 into ll.rtFreeList->dummy WRONG! */
 >           mr 3,10          /* put ll.rtFreeList into r3 (but should be 
 > the lnk
 >                                  pointer) */
 >           lwz 0,4(9)       /* load ll.rtFreeList->next into r0 */
 >           li 9,0
 >           stw 0,8(11)     /* store ll.rtFreeList->next into ll.rtFreeList */
 >           li 0,18
 >           stw 9,4(10)
 >           stb 0,42(10)    /* store 0x12 into space outside allocated data! */
 >           blr
 >  .Lfe1:
 >           .size   force_gcc_ppc_bug,.Lfe1-force_gcc_ppc_bug
 >           .comm   ll,12,4
 >           .ident  "GCC: (GNU) 3.1.1 20020704 (prerelease)"
 
 Ah, now I understand your complaint, but this is not a compiler bug. If you 
 look at your code
 
 static
 a_link *force_gcc_ppc_bug(void)
 {
    FreeList *freeNode;
    a_link    *lnk;
 
     /* pop one node */
     freeNode = ll.rtFreeList;
     ll.rtFreeList = ll.rtFreeList->next;
     freeNode->next = 0;
 
     lnk    = (a_link *) freeNode;
     lnk->a = 0x1234;
     lnk->l = 0x12;
 
     return lnk;
 }
 
 and check against the aliasing rules in the C standard, you'll notice that 
 you violate these rules thus invoking undefined behaviour. According to the 
 C standard, the compiler can assume that lnk and FreeNode point to 2 
 _distinct_ objects and optimizes accordingly.
 Either fixup your code or use -fno-strict-aliasing to turn off this 
 optimization at the price of slower code.
 
 I'll close the bug report.
 
 Franz.