[llvm] [RISCV] Add CompressPat for all cases in isCopyInstrImpl (PR #136875)

[Alex Bradbury via llvm-commits]

================
@@ -1053,3 +1057,21 @@ let Predicates = [HasStdExtCOrZca, IsRV64] in {
 def : CompressPat<(SD GPR:$rs2, SPMem:$rs1, uimm9_lsb000:$imm),
                   (C_SDSP GPR:$rs2, SPMem:$rs1, uimm9_lsb000:$imm)>;
 } // Predicates = [HasStdExtCOrZca, IsRV64]
+
+let isCompressOnly = true, Predicates = [HasStdExtCOrZca, HasStdExtZba] in {
+def : CompressPat<(SH1ADD GPRNoX0:$rs1, X0, GPRNoX0:$rs2),
----------------
asb wrote:

This is the kind trade-off I wondered about. I oversimplified a bit in my summary.

On the one hand: Although canonicalising earlier is theoretically better for this reason, it also seem defensible to optimise for existing real-world and likely future hardware where unless something big changes, it's hard to imagine it doesn't have RVC support. Especially as this is more of a micro-optimisation that we're mainly motivated to do to keep output clean and non-surprising for anyone doing perf analysis.

On the other hand: there's an attraction and probably minor engineering advantage to keeping differences between compressed/non-compressed as minimal as possible. And if non-C hardware did appear, our future selves would be thankful.

Overall I'm still torn...

https://github.com/llvm/llvm-project/pull/136875