[PATCH PR94442] [AArch64] Redundant ldp/stp instructions emitted at -O3

[xiezhiheng]

> -----Original Message-----
> From: Richard Sandiford [mailto:richard.sandiford@arm.com]
> Sent: Monday, July 6, 2020 5:31 PM
> To: xiezhiheng <xiezhiheng@huawei.com>
> Cc: Richard Biener <richard.guenther@gmail.com>; gcc-patches@gcc.gnu.org
> Subject: Re: [PATCH PR94442] [AArch64] Redundant ldp/stp instructions
> emitted at -O3
> 
> No, this is unfortunately a known bug.  See:
> 
>   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95964
> 
> (Although the PR is recent, it's been a known bug for longer.)
> 
> As you say, the difficulty is that the correct attributes depend on what
> the built-in function does.  Most integer arithmetic is “const”, but things
> get more complicated for floating-point arithmetic.
> 
> The SVE intrinsics use a three stage process:
> 
> - each function is classified into one of several groups
> - each group has a set of flags that describe what functions in the
>   group can do
> - these flags get converted into attributes based on the current
>   command-line options
> 
> I guess we should have something similar for the arm_neon.h built-ins.
> 
> If you're willing to help fix this, that'd be great.  I think a first
> step would be to agree a design.
> 
> Thanks,
> Richard

I'd like to have a try.  I have checked the steps in SVE intrinsics.
It defines a base class "function_base" and derives different classes
to describe several intrinsics for each.  And each class may
have its own unique flags described in virtual function "call_properties".
The specific attributes will be converted from these flags in
"get_attributes" later.

I find that there are more than 100 classes in total and if I only
need to classify them into different groups by attributes, maybe
we does not need so many classes?

The difficult thing I think is how to classify neon intrinsics into
different groups.  I'm going to follow up the way in SVE intrinsics
first now.

Xie Zhiheng