[DAG] combineShiftToAVG - failure to match avgceil patterns on preferIncOfAddToSubOfNot targets

[RKSimon]

define <vscale x 2 x i64> @rhadds_v2i64(<vscale x 2 x i64> %s0, <vscale x 2 x i64> %s1) {
; SVE-LABEL: rhadds_v2i64:
; SVE:       // %bb.0: // %entry
; SVE-NEXT:    eor z2.d, z0.d, z1.d
; SVE-NEXT:    orr z0.d, z0.d, z1.d
; SVE-NEXT:    asr z1.d, z2.d, #1
; SVE-NEXT:    sub z0.d, z0.d, z1.d
; SVE-NEXT:    ret
;
; SVE2-LABEL: rhadds_v2i64:
; SVE2:       // %bb.0: // %entry
; SVE2-NEXT:    ptrue p0.d
; SVE2-NEXT:    srhadd z0.d, p0/m, z0.d, z1.d
; SVE2-NEXT:    ret
  %s0s = sext <vscale x 2 x i64> %s0 to <vscale x 2 x i128>
  %s1s = sext <vscale x 2 x i64> %s1 to <vscale x 2 x i128>
  %add = add <vscale x 2 x i128> %s0s, splat (i128 1)
  %add2 = add <vscale x 2 x i128> %add, %s1s
  %s = ashr <vscale x 2 x i128> %add2, splat (i128 1)
  %result = trunc <vscale x 2 x i128> %s to <vscale x 2 x i64>
  ret <vscale x 2 x i64> %result
}

folded to:

define <vscale x 2 x i64> @rhadds_v2i64(<vscale x 2 x i64> %s0, <vscale x 2 x i64> %s1) {
  %s0s = sext <vscale x 2 x i64> %s0 to <vscale x 2 x i128>
  %s1s = sext <vscale x 2 x i64> %s1 to <vscale x 2 x i128>
  %not = xor <vscale x 2 x i128> %s0s, splat (i128 -1)
  %sub = sub <vscale x 2 x i128> %s1s, %not
  %s = ashr <vscale x 2 x i128> %sub, splat (i128 1)
  %result = trunc <vscale x 2 x i128> %s to <vscale x 2 x i64>
  ret <vscale x 2 x i64> %result
}

Targets such as aarch64 will fold add (add x, y), 1 to sub y, (xor x, -1), preventing the combineShiftToAVG AVGCEIL matching code from identifying the new pattern. This is a particular problem for the SVE/SVE2 tests as the above code will no longer compile.

Noticed while investigating regression for DAG topological sorting - which will fold the add to sub y, (xor x, -1) earlier than usual.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Simon Pilgrim (RKSimon)

```ll define <vscale x 2 x i64> @rhadds_v2i64(<vscale x 2 x i64> %s0, <vscale x 2 x i64> %s1) { ; SVE-LABEL: rhadds_v2i64: ; SVE: // %bb.0: // %entry ; SVE-NEXT: eor z2.d, z0.d, z1.d ; SVE-NEXT: orr z0.d, z0.d, z1.d ; SVE-NEXT: asr z1.d, z2.d, #1 ; SVE-NEXT: sub z0.d, z0.d, z1.d ; SVE-NEXT: ret ; ; SVE2-LABEL: rhadds_v2i64: ; SVE2: // %bb.0: // %entry ; SVE2-NEXT: ptrue p0.d ; SVE2-NEXT: srhadd z0.d, p0/m, z0.d, z1.d ; SVE2-NEXT: ret %s0s = sext <vscale x 2 x i64> %s0 to <vscale x 2 x i128> %s1s = sext <vscale x 2 x i64> %s1 to <vscale x 2 x i128> %add = add <vscale x 2 x i128> %s0s, splat (i128 1) %add2 = add <vscale x 2 x i128> %add, %s1s %s = ashr <vscale x 2 x i128> %add2, splat (i128 1) %result = trunc <vscale x 2 x i128> %s to <vscale x 2 x i64> ret <vscale x 2 x i64> %result } ``` folded to: ```ll define <vscale x 2 x i64> @rhadds_v2i64(<vscale x 2 x i64> %s0, <vscale x 2 x i64> %s1) { %s0s = sext <vscale x 2 x i64> %s0 to <vscale x 2 x i128> %s1s = sext <vscale x 2 x i64> %s1 to <vscale x 2 x i128> %not = xor <vscale x 2 x i128> %s0s, splat (i128 -1) %sub = sub <vscale x 2 x i128> %not, %s1s %s = ashr <vscale x 2 x i128> %sub, splat (i128 1) %result = trunc <vscale x 2 x i128> %s to <vscale x 2 x i64> ret <vscale x 2 x i64> %result } ```

Targets such as aarch64 will fold add (add x, y), 1 to sub y, (xor x, -1), preventing the combineShiftToAVG AVGFLOOR/CEIL matching code from identifying the new pattern. This is a particular problem for the SVE/SVE2 tests as the above code will no longer compile.

Noticed while investigating regression for DAG topological sorting - which will fold the add to sub y, (xor x, -1) earlier than usual.

[laurenmchin]

I'd like to take on this issue, please - could you assign this to me?

[RKSimon]

@laurenmchin DAGCombiner::foldShiftToAvg might be an easier place to add the SUB(X,NOT(Y)) AVGCEIL patterns

[laurenmchin]

@laurenmchin DAGCombiner::foldShiftToAvg might be an easier place to add the SUB(X,NOT(Y)) AVGCEIL patterns

great, thank you for the helpful suggestion