[DAG]SimplifyDemandedVectorElts-add ISD::AVGCEILS/AVGCEILU/AVGFLOORS/AVGFLOORU nodes

[aniplcc]

Fixes #84768

[llvmbot]

@llvm/pr-subscribers-backend-x86
@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-aarch64

Author: aniplcc (aniplcc)

Changes

Fixes #84768

---

Full diff: https://github.com/llvm/llvm-project/pull/86284.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+4)
• (modified) llvm/test/CodeGen/AArch64/hadd-combine.ll (+52)

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index da29b1d5b312f8..58d9394feb1237 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -3524,6 +3524,10 @@ bool TargetLowering::SimplifyDemandedVectorElts(
     }
     [[fallthrough]];
   }
+  case ISD::AVGCEILS:
+  case ISD::AVGCEILU:
+  case ISD::AVGFLOORS:
+  case ISD::AVGFLOORU:
   case ISD::OR:
   case ISD::XOR:
   case ISD::SUB:
diff --git a/llvm/test/CodeGen/AArch64/hadd-combine.ll b/llvm/test/CodeGen/AArch64/hadd-combine.ll
index e12502980790da..7c1c089839edc5 100644
--- a/llvm/test/CodeGen/AArch64/hadd-combine.ll
+++ b/llvm/test/CodeGen/AArch64/hadd-combine.ll
@@ -879,6 +879,58 @@ define <8 x i16> @uhadd_fixedwidth_v4i32(<8 x i16> %a0, <8 x i16> %a1)  {
   ret <8 x i16> %res
 }
 
+define <8 x i16> @shadd_demandedelts(<8 x i16> %a0, <8 x i16> %a1) {
+; CHECK-LABEL: shadd_demandedelts:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    shadd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    ret
+  %s0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> zeroinitializer
+  %op = call <8 x i16> @llvm.aarch64.neon.shadd.v8i16(<8 x i16> %s0, <8 x i16> %a1)
+  %r0 = shufflevector <8 x i16> %op, <8 x i16> undef, <8 x i32> zeroinitializer
+  ret <8 x i16> %r0
+}
+
+define <8 x i16> @srhadd_demandedelts(<8 x i16> %a0, <8 x i16> %a1) {
+; CHECK-LABEL: srhadd_demandedelts:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    srhadd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    ret
+  %s0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> zeroinitializer
+  %op = call <8 x i16> @llvm.aarch64.neon.srhadd.v8i16(<8 x i16> %s0, <8 x i16> %a1)
+  %r0 = shufflevector <8 x i16> %op, <8 x i16> undef, <8 x i32> zeroinitializer
+  ret <8 x i16> %r0
+}
+
+define <8 x i16> @uhadd_demandedelts(<8 x i16> %a0, <8 x i16> %a1) {
+; CHECK-LABEL: uhadd_demandedelts:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    uhadd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    ret
+  %s0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> zeroinitializer
+  %op = call <8 x i16> @llvm.aarch64.neon.uhadd.v8i16(<8 x i16> %s0, <8 x i16> %a1)
+  %r0 = shufflevector <8 x i16> %op, <8 x i16> undef, <8 x i32> zeroinitializer
+  ret <8 x i16> %r0
+}
+
+define <8 x i16> @urhadd_demandedelts(<8 x i16> %a0, <8 x i16> %a1) {
+; CHECK-LABEL: urhadd_demandedelts:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    urhadd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    dup v0.8h, v0.h[0]
+; CHECK-NEXT:    ret
+  %s0 = shufflevector <8 x i16> %a0, <8 x i16> undef, <8 x i32> zeroinitializer
+  %op = call <8 x i16> @llvm.aarch64.neon.urhadd.v8i16(<8 x i16> %s0, <8 x i16> %a1)
+  %r0 = shufflevector <8 x i16> %op, <8 x i16> undef, <8 x i32> zeroinitializer
+  ret <8 x i16> %r0
+}
+
 declare <8 x i8> @llvm.aarch64.neon.shadd.v8i8(<8 x i8>, <8 x i8>)
 declare <4 x i16> @llvm.aarch64.neon.shadd.v4i16(<4 x i16>, <4 x i16>)
 declare <2 x i32> @llvm.aarch64.neon.shadd.v2i32(<2 x i32>, <2 x i32>)

[RKSimon]

These don't appear to be working? I'd expect:

; CHECK:       // %bb.0:
; CHECK-NEXT:    shadd v0.8h, v0.8h, v1.8h
; CHECK-NEXT:    dup v0.8h, v0.h[0]
; CHECK-NEXT:    ret

[aniplcc]

my bad, forgot to run it against build, will update

[aniplcc]

Seems like the addition of the cases don't seem to be doing anything. I don't see any change in the resulting assembly.

[RKSimon]

If I had to guess the problem is that we don't hit this until after lowering and the AArch64 DUP node combines aren't calling SimplifyDemandedVectorElts - @davemgreen does that make sense to you?

[aniplcc]

Also, the ISD:: opcode passed to SimplifyDemandedVectorElts is INTRINSIC_WO_CHAIN(45) rather than one of AVGCEIL(S/U)/AVGFLOOR(S/U)?

[RKSimon]

Also, the ISD:: opcode passed to SimplifyDemandedVectorElts is INTRINSIC_WO_CHAIN(45) rather than one of AVGCEIL(S/U)/AVGFLOOR(S/U)?

That's prior to lowering - after legalization the intrinsic call will lower to a ISD::AVG* node, but by that point the shuffles have become AArch64ISD::DUPLANE* nodes.

[RKSimon]

@aniplcc Please can you rebase after 5d3ef06 and regenerate the test checks?

[RKSimon]

LGTM (keep the AArch64 tests - I'll create a followup issue to improve DUPLANE demanded elts handling)