Re: [PATCH] rs6000: replace '(const_int 0)' to 'unspec:BLK [(const_int 0)]' for stack_tie

[Richard Biener <richard.guenther@gmail.com>]

> Am 14.06.2023 um 17:41 schrieb Segher Boessenkool <segher@kernel.crashing.org>:
> 
> Hi!
> 
>> On Wed, Jun 14, 2023 at 07:59:04AM +0000, Richard Biener wrote:
>>> On Wed, 14 Jun 2023, Jiufu Guo wrote:
>>> 3. "set (mem/c:DI (reg/f:DI 1 1) unspec:DI (const_int 0 [0])
>>> UNSPEC_TIE".
>>>   This avoids using BLK on unspec, but using DI.
>> 
>> That gives the MEM a size which means we can interpret the (set ..)
>> as killing a specific area of memory, enabling DSE of earlier
>> stores.
> 
> Or DSE can delete this tie even, if it can see some later store to the
> same location without anything in between that can read what the tie
> stores.
> 
> BLKmode avoids all of this.  You can call that elegant, you can call it
> cheating, you can call it many things -- but it *works*.
> 
>> AFAIU this special instruction is only supposed to prevent
>> code motion (of stack memory accesses?) across this instruction?
> 
> Form rs6000.md:
> ; This is to explain that changes to the stack pointer should
> ; not be moved over loads from or stores to stack memory.
> (define_insn "stack_tie"

That suggests it’s the hard register value that‘s protected, not the memory pointed to.  I suppose that means an unspec volatile with the reg as input would serve the same?

Or maybe that’s not the whole story.


> and from rs6000-logue.cc:
> /* This ties together stack memory (MEM with an alias set of frame_alias_set)
>   and the change to the stack pointer.  */
> static void
> rs6000_emit_stack_tie (rtx fp, bool hard_frame_needed)

I cannot make sense of that comment, but not sure if I really want to know …

> A big reason this is needed is because of all the hard frame pointer
> stuff, which the generic parts of GCC require, but there is no register
> for that in the Power architecture.  Nothing is an issue here in most
> cases, but sometimes we need to do unusual things to the stack, say for
> alloca.
> 
>> I'd say a
>> 
>>  (may_clobber (mem:BLK (reg:DI 1 1)))
> 
> "clobber" always means "may clobber".  (clobber X) means X is written
> with some unspecified value, which may well be whatever value it
> currently holds.  Via some magical means or whatever, there is no
> mechanism specified, just the effects :-)
> 
>> might be more to the point?  I've used "may_clobber" which doesn't
>> exist since I'm not sure whether a clobber is considered a kill.
>> The docs say "Represents the storing or possible storing of an 
>> unpredictable..." - what is it?  Storing or possible storing?
> 
> It is the same thing.  "clobber" means the same thing as "set", except
> the value that is written is not specified.
> 
>> I suppose stack_tie should be less strict than the documented
>> (clobber (mem:BLK (const_int 0))) (clobber all memory).
> 
> "clobber" is nicer than the set to (const_int 0).  Does it work though?
> All this code is always fragile :-/  I'm all for this change, don't get
> me wrong, but preferably things stay in working order.
> 
> We use "stack_tie" as a last resort heavy hammer anyway, in all normal
> cases we explain the actual data flow explicitly and correctly, also
> between the various registers used in the *logues.
> 
> 
> Segher