On Tue, May 2, 2017 at 7:06 PM, Bin.Cheng <amker.ch...@gmail.com> wrote:
> On Mon, Apr 24, 2017 at 11:34 AM, Richard Biener
> <richard.guent...@gmail.com> wrote:
>> On Tue, Apr 18, 2017 at 12:41 PM, Bin Cheng <bin.ch...@arm.com> wrote:
>>> Hi,
>>> For now, we check validity of offset by computing the maximum offset then
>>> checking if
>>> offset is smaller than the max offset. This is inaccurate, for example,
>>> some targets
>>> may require offset to be aligned by power of 2. This patch introduces new
>>> interface
>>> checking validity of offset. It also buffers rtx among different calls.
>>>
>>> Is it OK?
>>
>> - static vec<HOST_WIDE_INT> max_offset_list;
>> -
>> + auto_vec<rtx> addr_list;
>> as = TYPE_ADDR_SPACE (TREE_TYPE (use->iv->base));
>> mem_mode = TYPE_MODE (TREE_TYPE (*use->op_p));
>>
>> - num = max_offset_list.length ();
>> + num = addr_list.length ();
>> list_index = (unsigned) as * MAX_MACHINE_MODE + (unsigned) mem_mode;
>> if (list_index >= num)
>>
>> num here is always zero and thus the compare is always true.
>>
>> + addr_list.safe_grow_cleared (list_index + MAX_MACHINE_MODE);
>> + for (; num < addr_list.length (); num++)
>> + addr_list[num] = NULL;
>>
>> the loop is now redundant (safe_grow_cleared)
>>
>> + addr = addr_list[list_index];
>> + if (!addr)
>> {
>>
>> always true again...
>>
>> I wonder if you really indented to drop 'static' from addr_list?
>> There's no caching
>> across function calls.
> Right, the redundancy is because I tried to cache across function
> calls with declarations like:
> static unsigned num = 0;
> static GTY ((skip)) rtx *addr_list = NULL;
> But this doesn't work, the addr_list[list_index] still gets corrupted somehow.
Well, you need GTY (()), not GTY((skip)) on them. Not sure if it works
for function-scope decls, you have to check. Look at whether a GC
root is created for the variable in gt-tree-ssa-loop-ivopts.h (need to tweak
GTFILES in the makefile plus include that generated file). tree-ssa-address.c
uses a global root for mem_addr_template_list for example.
Richard.
> Thanks,
> bin
>>
>> + /* Split group if aksed to, or the offset against the first
>> + use can't fit in offset part of addressing mode. IV uses
>> + having the same offset are still kept in one group. */
>> + if (offset != 0 &&
>> + (split_p || !addr_offset_valid_p (use, offset)))
>>
>> && goes to the next line.
>>
>> Richard.
>>
>>
>>
>>> Thanks,
>>> bin
>>> 2017-04-11 Bin Cheng <bin.ch...@arm.com>
>>>
>>> * tree-ssa-loop-ivopts.c (compute_max_addr_offset): Delete.
>>> (addr_offset_valid_p): New function.
>>> (split_address_groups): Check offset validity with above function.
