On Thu, Dec 2, 2021 at 12:37 AM Qing Zhao <qing.z...@oracle.com> wrote:
>
> Hi,
>
> As I studied, the issue with this bug is similar as PR102285:
>
> For the testing case:
>
> int
> qy (void)
> {
> int tw = 4;
> int fb[tw];
> return fb[2]; /* { dg-warning "uninitialized" } */
> }
>
> if compiled with
> /home/opc/Install/cross/bin/riscv64-unknown-linux-gnu-gcc -c -O1
> -mno-strict-align -ftrivial-auto-var-init=pattern -fdump-tree-all t.c
>
> ******t.c.032t.cpp1:
> int qy ()
> {
> unsigned char fb.3[16];
> int tw;
> int[0:D.1494] * fb.1;
> sizetype D.1494;
> void * saved_stack.2_16;
> int _27;
>
> <bb 2> :
> saved_stack.2_16 = __builtin_stack_save ();
> fb.1_25 = &fb.3;
> (*fb.1_25) = .DEFERRED_INIT (16, 1, 1);
> _27 = (*fb.1_25)[2];
> fb.3 ={v} {CLOBBER};
> __builtin_stack_restore (saved_stack.2_16);
> return _27;
>
> }
>
> ******t.c.033t.forwprop1:
> int qy ()
> {
> unsigned char fb.3[16];
> int tw;
> int[0:D.1494] * fb.1;
> sizetype D.1494;
> void * saved_stack.2_16;
> int _27;
>
> <bb 2> :
> saved_stack.2_16 = __builtin_stack_save ();
> MEM[(int[0:D.1494] *)&fb.3] = .DEFERRED_INIT (16, 1, 1);
> _27 = MEM[(int[0:D.1494] *)&fb.3][2];
> fb.3 ={v} {CLOBBER};
> __builtin_stack_restore (saved_stack.2_16);
> return _27;
>
> }
>
> the problem with the above IR for .DEFERRED_INIT expansion is:
>
> for MEM[(int[0:D.1494] *)&fb.3] = .DEFERRED_INIT (16, 1, 1)
>
> after the “forwprop1” phase, the original address that point to VLA fb.1 is
> replaced with the the new address that point to fixed array fb.3, however,
> the TREE_TYPE of the MEM still is kept the old VLA type int[0:D.1984]. Such
> inconsistency causes the ICE during expanding of call to the above
> .DEFERRED_INIT.
>
> If during the forwprop, when replacing “fb.1” with “&fb.3”, the TREE_TYPE of
> the MEM_REF can be updated accordingly, we should fix this inconsistency of
> the IR, and therefore fix the issue in DEFERRED_INIT expansion.
>
> So, I think that the fix should be in “forwprop”, and update the TREE_TYPE of
> the MEM_REF accordingly.
Actually the IL is kept consistent by preserving the original
TREE_TYPE of the MEM_REF. The issue is that deferred_init
expansion sticks to the original type rather than initializing the
underlying storage.
See my comments in the PR.
> What do you think on this?
>
> thanks.
>
> Qing
>
>
