search

llvm / llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies.
http://llvm.org
Other
28.69k stars 11.86k forks source link

[X86] Better scheduling/selection of the LEA instruction #31673

Open RKSimon opened 7 years ago

RKSimon commented 7 years ago
Bugzilla Link 32326
Version trunk
OS Windows NT
Blocks llvm/llvm-project#31672
CC @adibiagio,@chriselrod,@gchatelet,@hfinkel,@rotateright,@ZviRackover

Extended Description

Some basic forms of the LEA instruction (2 source operands, no scale etc.) can typically be performed on the cpu's generic ALUs whilst the complex forms (3 source operands, scale + offset etc.) can only be performed on a cpus's AGUs.

We need to better tag the different LEA instructions so that we can discriminate in the scheduler model and compare them against other memory address instructions.

We should then be able to improve LEA pattern selection in the machine combiner (balance ALU/AGU usage, use multi stage LEAs for simple integer multiplies etc.).

adibiagio commented 5 years ago

Ping - Hopefully llvm-mca + variants can help us sort this one out and llvm-exegesis can test it.

This has been fixed for BtVer2 only, at revision 337469 http://llvm.org/viewvc/llvm-project?view=revision&revision=337469

It should be easy to use schedule 'write variant' definitions to model the latency/throughput of LEA on other X86 processors.

RKSimon commented 6 years ago

Ping - Hopefully llvm-mca + variants can help us sort this one out and llvm-exegesis can test it.