This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [ia64] [PR target/57491] internal compiler error: in ia64_split_tmode -O2, quadmath


> As far as I understand semantics of this insn:
>   (insn 200 199 0 (set (reg:DI 15 r15)
>       (mem:DI (post_dec:DI (reg/f:DI 15 r15 [447])) [3
>       *_61[_12]{lb: 1 sz: 64}.text+8 S8 A64])) -1 (nil))
> What is done is (in that sequence).
>   1. Calculate address of MEM: get r15 value.
>   2. Decrement r15 value.
>   3. Load MEM in to r15.
> 
> Point 2 is useless as we kill it by 3.
> So, it is clobbered and as mention in comment this is sometimes ok to
> override pointer with pointer value.

That depends on the semantics of the hardware instruction though, does it 
really guarantee 1/2/3 in that order?

> We need to set `dead' flag only when address is actually going to be killed
> by load.
> 
> Patch in the bottom. Test from PR pass.

The patch looks good to me if you also adjust the last sentence in the comment 
just above the block:

  /* It is possible for reload to decide to overwrite a pointer with
     the value it points to.  In that case we have to do the loads in
     the appropriate order so that the pointer is not destroyed too
     early.  Also we must not generate a postmodify for that second
     load, or rws_access_regno will die.  */

Something like "And we must not generate a postmodify for the second load if
the destination register overlaps with the base register".

Thanks for fixing this.

-- 
Eric Botcazou


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]