[patch,avr] PR87376: Disable -ftree-ter

[Jeff Law]

On 7/3/24 1:26 PM, Georg-Johann Lay wrote:
> 
> 
> Am 02.07.24 um 15:48 schrieb Richard Biener:
>> On Tue, Jul 2, 2024 at 3:43 PM Georg-Johann Lay <avr@gjlay.de> wrote:
>>>
>>> Hi Jeff,
>>>
>>> This is a patch to get correct code out of 64-bit
>>> loads from address-space __memx.
>>>
>>> The AVR address-spaces may require that move insns issue
>>> calls to library support functions, a fact that -ftree-ter
>>> doesn't account for.  tree-ssa-ter.cc then replaces an
>>> expression across such a library call, resulting in wrong code.
>>>
>>> This patch disables that pass per default on avr, as there is no
>>> more fine grained way to avoid malicious optimizations.
>>> The pass can still be re-enabled by means of explicit -ftree-ter.
>>>
>>> Ok to apply?
>>
>> I think this requires more details on what goes wrong - I assume
>> it's not stmt reordering that effectively happens but recursive
>> expand_expr on SSA defs when those invoke libcalls?  In that
>> case this would point to a deeper issue.
> 
> The difference is that with TER, we get a hard reg in .expand
> for a movdi from 24-bit address-space __memx.
> 
> Such moves require library calls, which in turn require
> specific hard registers.  As avr backend has no movdi, the
> moddi gets expanded as 8 * movqi, and that does not work
> when the target registers are hard regs, as some of them
> are clobbered by the libcalls.
But this is something that's handled for other targets.  I don't 
remember all the details, but there's generic code to handle this situation.

Jeff