[llvm-dev] [X86] Memory folding tables in x86 backend

[Chris Lattner via llvm-dev]

> On Jul 23, 2017, at 7:09 AM, Simon Pilgrim via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> 
>> On 23 Jul 2017, at 12:19, Joerg Sonnenberger via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>> 
>> On Sun, Jul 23, 2017 at 08:48:00AM +0000, Musa, Ayman via llvm-dev wrote:
>>> 3-     Give up on the auto-generation idea and manually update the current tables iteratively with new chunks of instructions until full state is achieved.
>>> 
>>> P.s. The TableGen backend emits more than 5200 entries, while the known exception at this point are ~20.
>> 
>> Why can't you do a variation of this by tagging the instructions with a
>> flag? I.e. for the rm variant, add a flag that says "this has an
>> equivalent register operand version". Given that a lot of the instruction
>> patterns are created via multi-classes, I would expect that to require a
>> lot less than 5200 updates.
> 
> +1 An opt-in ‘GenerateMemoryFolding’ style flag (defaulting to false) seems safer to me than than trying to ensure you’ve correctly flagged all illegal instruction fold.

The counterargument to this is that it adds a lot of cruft to the td files.  It should be a lot more rare for an instruction to be “unfoldable” than to be “foldable”.

Validation should be straight-foward for this: generate the table automatically using the new tool and see what the differences are.  If the differences are due to incorrect additions, just add the flag to them.

BTW, adding a tblgen backend to automate this is totally awesome.  Thank you for pushing on this Ayman!

-Chris