RFR: JDK-8323497: On x64, use 32-bit immediate moves for narrow klass base if possible

[Thomas Stuefe]

On Wed, 10 Jan 2024 18:44:57 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> @shipilev It is just the code representation of the matching, the corresponding nodes are `loadConUL32`, `loadConL32` and `loadConL` and the matcher uses node costs to sort out the priority.
>> 
>> https://github.com/openjdk/jdk/blob/c1282b57f50002edd08c93aed784390cca83b9b8/src/hotspot/cpu/x86/x86_64.ad#L4807
>
> Oh, subtle. Thanks!

With the new movptr variant, things look good:

< 2g


 ;; decode_klass_not_null
  0x00007fcb608681c4:   mov    $0x18000000,%r11d
0x00007fcb608681c4:   41 bb 00 00 00 18 
  0x00007fcb608681ca:   add    %r11,%r10
0x00007fcb608681ca:   4d 03 d3 


< 4g


 ;; decode_klass_not_null
  0x00007f382c8681c4:   mov    $0x85000000,%r11d
0x00007f382c8681c4:   41 bb 00 00 00 85 
  0x00007f382c8681ca:   add    %r11,%r10
0x00007f382c8681ca:   4d 03 d3 


> 4g


 ;; decode_klass_not_null
  0x00007fd8908681c4:   movabs $0x7fd7b6000000,%r11
0x00007fd8908681c4:   49 bb 00 00 00 b6 d7 7f 00 00 
  0x00007fd8908681ce:   add    %r11,%r10
0x00007fd8908681ce:   4d 03 d3 


I assume if the 32-bit mov variants were to use non-Rxx registers, the REX prefix could also be omitted? Giving us another byte back.

I'll run tests with the new variant.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/17340#discussion_r1448706776