[llvm-dev] GlobalISel design update and goals

[Andrew Haley via llvm-dev]

On 08/03/2018 04:56 AM, Leslie Zhai wrote:
> * There is no instruction selection "concept" equivalent to LLVM's 
> SelectionDAG and GlobalISel in HotSpot C1? Because I manually write 
> assembly[1] lowing HIR to LIR in HotSpot C1. So which one is better? 
> LLVM or HotSpot selection by human?

The goal of C1 is to be simple and fast.  C2 does the pattern-matching
instruction generation.

> * Why not use Greedy, just like LLVM's RegAllocGreedy, to take place of 
> Linear Scan[2] for HotSpot C1's register allocation?

Same reason as above.  Also, there will be very little popularity for major
rewrites of C1.  It is what it is.

-- 
Andrew Haley
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671