Re: [PATCH] fwprop: Allow UNARY_P and check register pressure.

[Jeff Law <jeffreyalaw@gmail.com>]

On 8/24/23 08:06, Robin Dapp via Gcc-patches wrote:
> Ping.  I refined the code and some comments a bit and added a test
> case.
> 
> My question in general would still be:  Is this something we want
> given that we potentially move some of combine's work a bit towards
> the front of the RTL pipeline?
> 
> Regards
>   Robin
> 
> Subject: [PATCH] fwprop: Allow UNARY_P and check register pressure.
> 
> This patch enables the forwarding of UNARY_P sources.  As this
> involves potentially replacing a vector register with a scalar register
> the ira_hoist_pressure machinery is used to calculate the change in
> register pressure.  If the propagation would increase the pressure
> beyond the number of hard regs, we don't perform it.
> 
> gcc/ChangeLog:
> 
> 	* fwprop.cc (fwprop_propagation::profitable_p): Add unary
> 	handling.
> 	(fwprop_propagation::update_register_pressure): New function.
> 	(fwprop_propagation::register_pressure_high_p): New function
> 	(reg_single_def_for_src_p): Look through unary expressions.
> 	(try_fwprop_subst_pattern): Check register pressure.
> 	(forward_propagate_into): Call new function.
> 	(fwprop_init): Init register pressure.
> 	(fwprop_done): Clean up register pressure.
> 	(fwprop_insn): Add comment.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* gcc.target/riscv/rvv/autovec/binop/vadd-vx-fwprop.c: New test.
So I was hoping that Richard S. would chime in here as he knows this 
code better than anyone.

This looks like a much better implementation of something I've done 
before :-)  Basically imagine a target where a sign/zero extension can 
be folded into arithmetic for free.  We put in various hacks to this 
code to encourage more propagations of extensions.

I still think this is valuable.  As we lower from gimple->RTL we're 
going to still have artifacts in the RTL that we're going to want to 
optimize away.  fwprop has certain advantages over combine, including 
the fact that it runs earlier, pre-loop.


It looks generally sensible to me.  But give Richard S. another week to 
chime in.  He seems to be around, but may be slammed with stuff right now.

jeff