[PATCH] D149383: [SelectionDAG][WIP] Add support for evaluating SetCC based on knownbits

[Noah Goldstein via Phabricator via llvm-commits]

goldstein.w.n added inline comments.


================
Comment at: llvm/test/CodeGen/X86/fold-rmw-ops.ll:1359
+; CHECK-NEXT:    movb $1, %al # encoding: [0xb0,0x01]
+; CHECK-NEXT:    testb %al, %al # encoding: [0x84,0xc0]
 ; CHECK-NEXT:    jne b # TAILCALL
----------------
craig.topper wrote:
> goldstein.w.n wrote:
> > These `movb $1, %al; testb %al, %al`'s (here and in many other cases) are unnecessary. I assume its because SelectionDAG only has BB view, so even if we can rule out some BBs (based on known true/false br-cond), there is no pass for that. Is there anything we can/should do about that?
> > 
> > Also NB, we really should never emit `movb $1, %al; testb %al, %al` just grabbing any gpr (I guess least recently used to minimize potential latency) and do `cmpb %gpr8, gpr8` then `jne`/`je` depending if we want it to be always true/false.
> > 
> Hopefully most of these are just tests that should have been folded by InstCombine or other passes earlier and not really cases that originate in SelectionDAG.
Do you think these regressions are something to worry about? Or acceptable as cases we would never expect to get from the middle-end.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D149383/new/

https://reviews.llvm.org/D149383