Re: [PATCH] Check if loading const from mem is faster

[Jiufu Guo <guojiufu@linux.ibm.com>]

Segher Boessenkool <segher@kernel.crashing.org> writes:

> On Wed, Feb 23, 2022 at 02:02:59PM +0100, Richard Biener wrote:
>> I'm assuming we're always dealing with
>> 
>>   (set (reg:MODE ..) <src_folded>)
>> 
>> here and CSE is not substituting into random places of an
>> instruction(?).  I don't know what 'rtx_cost' should evaluate
>> to for a constant, if it should implicitely evaluate the cost
>> of putting the result into a register for example.
>
> rtx_cost is no good here (and in most places).  rtx_cost should be 0
> for anything that is used as input in a machine instruction -- but you
> need much more context to determine that.  insn_cost is much simpler and
> much easier to use.
>
>> Using RTX_COST with SET and 1 at least looks no worse than using
>> your proposed new target hook and comparing it with the original
>> unfolded src (again with SET and 1).
>
> It is required to generate valid instructions no matter what, before
> the pass has finished that is.  On all more modern architectures it is
> futile to think you can usefully consider the cost of an RTL expression
> and derive a real-world cost of the generated code from that.

Thanks Segher for pointing out these!  Here is  another reason that I
did not use rtx_cost: in a few passes, there are codes to check the
constants and store them in constant pool.  I'm thinking to integerate
those codes in a consistent way.


BR,
Jiufu

>
> But there is so much more wrong with cse.c :-(
>
>
> Segher