CT_00_031 -- Additional Constant Synthesis Improvements

Constant synthesis was greatly improved during 1H 2024.  The areas noted here are exploratory and thus some ideas may be useful, some may not.  These are primarily meant to improve pack sequences, though they may be helpful for designs without Zbkb.


Areas to explore:

  1. For a constant where the high and low halves only differ by a few bits, consider using pack as an intermediate step and the Zbs instructions to make final adjustments.
  2. When possible derive the upper and lower halves from a common sequence, so for example sharing a lui
    1. This can be handled during synthesis
    2. May also be capturable during reload_move2add
  3. Similarly for cases where the high half can be constructed by merely shifting the low half.
  4. Improve handling of constants where the upper and lower 32bit halves are the same and zbkb is not available.  It is believed this is a relatively small adjustment to the existing code.
  5. SImilar to the prior item with both halves equal, but when bit 31/63 are both on.  For this case generate ~val (which will necessarily have bits 31/63 off), then xor against -1 to get the final target value.
  6. Reevaluate the mvconst_internal pattern.   There's some belief that it could be eliminated or used much less often.  Right now it "lies" to the combiner and indicates it can load an arbitrary 64bit constant as a single instruction.  That's a useful property when we can to combine the constant in another instruction, but it's painful in other ways such as its interaction with the code to improve additional of constants that are not simm12, but are 2*simm12.
  7. Some constants may be generated with something like lui+shadd+addi where the lui+shadd targets a nearby constant and the final addi fixes things up.  The approach of targeting a nearby constant with a fixup at the end probably applies to most synthesis opportunities where the cost is 4+ instructions.
  8. Investigate reload_cse_move2add which should allow deriving some constants from other nearby constants.  In theory it should "just work", but we should verify and if it isn't fix it


Stakeholders/Partners

RISE:

Ventana: Jeff Law & Raphael Zinsly – general oversight / guidance and implementation work.

Rivos: Vineet Gupta – generally interested in constant synthesis and has some state on #5 above.

External:



Dependencies



Status

Development

IN PROGRESS


Development Timeline2H2024
Upstreaming

IN PROGRESS



Upstream Version

gcc-15 (target)

(Spring 2025)





Contacts

Jeff Law (Ventana)


DependenciesNone


Updates

 

  • Raphael's work for generating constants where the upper/low halves are the same with bit 31/63 on has been integrated upstream.

 

  • Matched upper/lower bits going through external review
  • Upper half derivable via shifting lower half also going through review

 

  • Noted as 2H 2024 work item