[PATCH] i386: Separate costs of pseudo registers from hard registers

[Jeff Law]

On 7/23/19 3:57 PM, H.J. Lu wrote:
[ Snip ]
> Here is the updated patch to improve register allocator and RTL
> expressions independently.
> 
> Any comments?
> 
> Thanks.
> 
> 
> -- H.J.
> 
> 
> 0001-i386-Separate-costs-of-pseudo-registers-from-hard-re.patch
> 
> From 79834daf252cecfc3ee51acd864641d2cdaff733 Mon Sep 17 00:00:00 2001
> From: "H.J. Lu" <hjl.tools@gmail.com>
> Date: Fri, 14 Jun 2019 13:30:16 -0700
> Subject: [PATCH] i386: Separate costs of pseudo registers from hard registers
> 
> processor_costs has costs of RTL expressions with pseudo registers and
> and costs of hard register moves:
> 
> 1. Costs of RTL expressions are used to generate the most efficient RTL
> operations with pseudo registers.
> 
> 2. Costs of hard register moves are used by register allocator to
> decide how to allocate and move hard registers.
> 
> Since relative costs of pseudo register load and store versus pseudo
> register moves in RTL expressions can be different from relative costs
> of hard registers, we should separate costs of RTL expressions with
> pseudo registers from costs of hard registers so that register allocator
> and RTL expressions can be improved independently.
> 
> This patch moves costs of hard register moves to the new hard_register
> field and duplicates costs of moves which are also used for costs of RTL
> expressions.
> 
> 	PR target/90878
> 	* config/i386/i386.c (inline_memory_move_cost): Use hard_register
> 	for costs of hard register moves.
> 	(ix86_register_move_cost): Likewise.
> 	* config/i386/i386.h (processor_costs): Move costs of hard
> 	register moves to hard_register.  Add int_load, int_store,
> 	xmm_move, ymm_move, zmm_move, sse_to_integer, integer_to_sse,
> 	sse_load, sse_store, sse_unaligned_load and sse_unaligned_store
> 	for costs of RTL expressions.
> 	* config/i386/x86-tune-costs.h: Move costs of hard register
> 	moves to hard_register.  Duplicate int_load, int_store,
> 	xmm_move, ymm_move, zmm_move, sse_to_integer, integer_to_sse,
> 	sse_load, sse_store for costs of RTL expressions.
This looks reasonable to me.  If you haven't had objections from Jan or
Uros, go ahead and commit it.

I'm assuming this patch isn't supposed to actually change anything yet
and a subsequent patch will twiddle some of the costs, particularly for
skylake.

jeff