[DAG] Enhance canCreateUndefOrPoison for VECTOR_COMPRESS and add tests for freeze elimination

Author: Michael-Chen-NJU

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -5705,7 +5705,9 @@ bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, const APInt &DemandedElts,
   case ISD::VECTOR_COMPRESS:
     // Return true only if undef is checked and at least one element is
     // demanded.
-    return !PoisonOnly && !DemandedElts.isZero();
+    if (Op.getOperand(2).isUndef())
+      return !PoisonOnly && !DemandedElts.isZero();
+    return false;
 
   default:
     // Allow the target to implement this method for its nodes.

llvm/test/CodeGen/X86/vector-compress-freeze.ll

@@ -0,0 +1,39 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+avx512f,+avx512vl | FileCheck %s
+
+declare <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32>, <16 x i1>, <16 x i32>)
+
+define <16 x i32> @test_compress_freeze_elimination(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a3) {
+; CHECK-LABEL: test_compress_freeze_elimination:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
+; CHECK-NEXT:    vpmovzxbd {{.*#+}} zmm1 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero,xmm2[4],zero,zero,zero,xmm2[5],zero,zero,zero,xmm2[6],zero,zero,zero,xmm2[7],zero,zero,zero,xmm2[8],zero,zero,zero,xmm2[9],zero,zero,zero,xmm2[10],zero,zero,zero,xmm2[11],zero,zero,zero,xmm2[12],zero,zero,zero,xmm2[13],zero,zero,zero,xmm2[14],zero,zero,zero,xmm2[15],zero,zero,zero
+; CHECK-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
+; CHECK-NEXT:    vpcompressd %zmm1, %zmm0 {%k1}
+; CHECK-NEXT:    retq
+  %cmp = icmp sgt <16 x i32> %a0, %a1
+  %ext = zext <16 x i8> %a3 to <16 x i32>
+  %cpr = call <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32> %ext, <16 x i1> %cmp, <16 x i32> splat(i32 15))
+  %fr = freeze <16 x i32> %cpr
+  %and = and <16 x i32> %fr, splat(i32 255)
+  ret <16 x i32> %and
+}
+
+; Test 2: Negative Case (Optimization must NOT happen)
+; PassThru is 'undef', so compress *can* generate undefs.
+; The freeze must remain to define those lanes.
+define <16 x i32> @test_compress_freeze_must_remain(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a3) {
+; CHECK-LABEL: test_compress_freeze_must_remain:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vpcmpgtd %zmm1, %zmm0, %k1
+; CHECK-NEXT:    vpmovzxbd {{.*#+}} zmm0 = xmm2[0],zero,zero,zero,xmm2[1],zero,zero,zero,xmm2[2],zero,zero,zero,xmm2[3],zero,zero,zero,xmm2[4],zero,zero,zero,xmm2[5],zero,zero,zero,xmm2[6],zero,zero,zero,xmm2[7],zero,zero,zero,xmm2[8],zero,zero,zero,xmm2[9],zero,zero,zero,xmm2[10],zero,zero,zero,xmm2[11],zero,zero,zero,xmm2[12],zero,zero,zero,xmm2[13],zero,zero,zero,xmm2[14],zero,zero,zero,xmm2[15],zero,zero,zero
+; CHECK-NEXT:    vpcompressd %zmm0, %zmm0 {%k1} {z}
+; CHECK-NEXT:    vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %zmm0
+; CHECK-NEXT:    retq
+  %cmp = icmp sgt <16 x i32> %a0, %a1
+  %ext = zext <16 x i8> %a3 to <16 x i32>
+  %cpr = call <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32> %ext, <16 x i1> %cmp, <16 x i32> undef)
+  %fr = freeze <16 x i32> %cpr
+  %and = and <16 x i32> %fr, splat(i32 255)
+  ret <16 x i32> %and
+}