On 10/31/2017 09:32 AM, Segher Boessenkool wrote:
> On Tue, Oct 31, 2017 at 09:12:25AM -0600, Jeff Law wrote:
>> On 10/31/2017 03:41 AM, Segher Boessenkool wrote:
>>> If the user asks for a stack clash probe interval of 64kB, we currently
>>> generate a "stdu rX,-65536(r1)" instruction. That instruction does not
>>> exist (the offset is a 16-bit signed number). If the offset is too big
>>> we should force it into a register and generate a "stdux rX,rY,r1"
>>> instruction, instead.
>>>
>>> Bootstrapped and regression checked on powerpc64-linux {-m32,-m64};
>>> committing to trunk.
>>>
>>>
>>> Segher
>>>
>>>
>>> 2017-10-31 Segher Boessenkool <seg...@kernel.crsahing.org>
>>>
>>> PR target/82674
>>> * config/rs6000/rs6000.md (allocate_stack): Force update interval
>>> into a register if it does not fit into an immediate offset field.
>> :-) That's one I had a fix for. But the issues with large probing
>> intervals are much more serious. In particular dealing with large
>> probing intervals for allocations in the prologue is a significantly
>> tougher problem due to the lack of a free hard reg.
>
> Yes, but I haven't seen testcases for that? The PR was just this.
:-) I just turned stack clash on by default, cranked up the probing
interval and ran the testsuite and started looking at the root cause of
the failures.
>
> We do have more than one register we can use here, FWIW. It gets a
> bit nasty because it depends on which ABI you're on exactly, which.
BIg sigh... There are times I wonder if I'd be happier as an app
developer or database junkie. Instead I chose to do low level stuff :(
jeff
