Implement machine dependent peephole optimizer at lower time

[russellhadley]

New pass is intended to exploit machine dependent instructions and allow for low level optimization based on specific instruction semantics.

Goals:

• Run before register allocation to remove issues with false dependencies and allow reduction in register pressure.
• Exploit specific features of the target ISA. Including but not limited to:
  • Immediate encoding size/shift semantics
  • Particular condition flag implementation
  • Target dependent address mode formation
• Allow more sophisticated instruction selection:
  • bt{s|r|c} formation
  • Aggressive optimization around condition flag/branch sequences.
  • Avoid machine glass jaws like LCP.
• Reorder/rework instructions to reduce register pressure.

Particular open questions:

• Which dataflow formulation should be used? Adjacent in window? Expression temp def/use? Unaliased SSA or some cheaper extended basic block approximation?
• Run just before RA or at higher tier run after?
• How to encapsulate a transform.
• How to enforce a high level of debug dump/tracing functionality for managing the transforms.

This is a big and perhaps controversial feature, but is intended to allow for more proactive engagement on code selection opts for collaborators with interests in particular targets.

category:design
theme:optimization
skill-level:expert
cost:extra-large
impact:large

[russellhadley]

Add random repro case for bit shifting that we're currently missing on x86. https://gist.github.com/russellhadley/e55fc9a918626166d98077aa3c8047c6

[mikedn]

Add random repro case for bit shifting that we're currently missing on x86.

Presumably you're referring to the and x, 31 added by the C# compiler. That can be easily be removed in lowering or morph. I once tried in morph, saves ~500 bytes in jitdiff fx.

[russellhadley]

@mikedn If you have the change we'd take it. :) The example was in an old issue in our internal database and I was just cleaning house.

[mikedn]

@russellhadley The morph version is in dotnet/coreclr#8744. I did it in morph thinking that it may be a good idea to get rid of unnecessary instructions early but then gave up on it because morph is already too hairy. I suppose I'll make a lower version one of these days. After all the example you have is written by me, I just found it on my HDD :)

[mikedn]

PR for shift count masking removal in lowering: dotnet/coreclr#11594