Re: [PATCH] RISC-V: Produce better code with complex constants [PR95632] [PR106602]

[Jeff Law <jeffreyalaw@gmail.com>]

On 12/7/22 13:55, Raphael Moreira Zinsly wrote:
> Due to RISC-V limitations on operations with big constants combine
> is failing to match such operations and is not being able to
> produce optimal code as it keeps splitting them. By pretending we
> can do those operations we can get more opportunities for
> simplification of surrounding instructions.
> 
> 2022-12-06 Raphael Moreira Zinsly <rzinsly@ventanamicro.com>
>             Jeff Law <jlaw@ventanamicro.com>
> 
> gcc/Changelog:
> 	PR target/95632
>          PR target/106602
>          * config/riscv/riscv.md: New pattern to simulate complex
>          const_int loads.
> 
> gcc/testsuite/ChangeLog:
> 	* gcc.target/riscv/pr95632.c: New test.
>          * gcc.target/riscv/pr106602.c: Likewise.
So to give a little background to others.

The core issue is that when we break down constants early, it can make 
it difficult for combine to reconstruct the constant and simplify code 
using the reconstructed constant -- you end up trying to do 4->3 or 
worse combination sequences which aren't supported by the combiner.

Usually this kind of scenario is handled with a "bridge" pattern.  Those 
are generally defined as patterns that exist solely for combine and may 
not correspond to any real instruction on the target.  "bridge" patterns 
are typically 2->1 or 3->1 combinations and are intermediate steps for 
4->N or even larger combination opportunities.  Obviously if the bridge 
doesn't allow subsequent simplifications, then the bridge pattern must 
generate correct code (either by generating suitable assembly or 
splitting later).

Raphael's patch introduces a bridge pattern that pretends we can load up 
splittable constants in a single insn.  We restrict the bridge pattern 
to be active from the point when CSE is no longer expected through the 
combiner up to the first splitter pass (where we'll break it down again 
if it's still in the IL).

So we get most of the benefit of splitting constants early (CSE, LICM, 
etc) while also getting the benefits of splitting late (combine 
simplifications).

Given I was working with Raphael on the patch, it's probably best for 
someone else to do the review rather than me approving it :-)

Jeff