Combine and(X, shuffle(X, pow 2 mask)) to all true

Combine N = and(X, shuffle_vector(X, power of 2 mask)) to all true.
Where X is either N or setcc(v, <0>, ne) or a bitcast of said setcc.

Author: badumbatish

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

@@ -18,12 +18,14 @@
 #include "WebAssemblySubtarget.h"
 #include "WebAssemblyTargetMachine.h"
 #include "WebAssemblyUtilities.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineJumpTableInfo.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SDPatternMatch.h"
 #include "llvm/CodeGen/SelectionDAG.h"
 #include "llvm/CodeGen/SelectionDAGNodes.h"
 #include "llvm/IR/DiagnosticInfo.h"
@@ -184,6 +186,10 @@ WebAssemblyTargetLowering::WebAssemblyTargetLowering(
     // Combine partial.reduce.add before legalization gets confused.
     setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
 
+    // Combine EXTRACT VECTOR ELT of AND(AND(X, SHUFFLE(X)), SHUFFLE(...)), 0
+    // to all_true
+    setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
+
     // Combine wide-vector muls, with extend inputs, to extmul_half.
     setTargetDAGCombine(ISD::MUL);
 
@@ -3287,6 +3293,85 @@ static SDValue performSETCCCombine(SDNode *N,
 
   return SDValue();
 }
+static SmallVector<int> buildMaskArrayByPower(int Power, size_t NumElements) {
+  // Generate 1-index array of elements from 2^Power to 2^(Power+1) exclusive
+  // The rest is filled with -1.
+  //
+  // For example, with NumElements = 4:
+  // When Power = 0: <1 -1 -1 -1>
+  // When Power = 1: <2  3 -1 -1>
+  // When Power = 2: <4  5  6  7>
+
+  uint From = pow(2, Power), To = pow(2, Power + 1);
+  assert(From < NumElements && To <= NumElements);
+
+  SmallVector<int> Res;
+  for (uint I = From; I < To; I++)
+    Res.push_back(I);
+  Res.resize(NumElements, -1);
+
+  return Res;
+}
+static SDValue matchAndOfShuffle(SDNode *N, int Power) {
+  // Matching on the case of
+  //
+  // Base case: A [bitcast for a] setcc(v, <0>, ne).
+  // Recursive case: N = and(X, shuffle(X, power mask)) where X is either
+  // recursive or base case.
+  using namespace llvm::SDPatternMatch;
+
+  EVT VT = N->getValueType(0);
+
+  SDValue LHS = N->getOperand(0);
+  int NumElements = VT.getVectorNumElements();
+  if (NumElements < pow(2, Power))
+    return SDValue();
+
+  if (N->getOpcode() != ISD::AND && NumElements == pow(2, Power)) {
+    SDValue BitCast, Matched;
+
+    // Try for a setcc first.
+    if (sd_match(N, m_c_SetCC(m_Value(Matched), m_Zero(),
+                              m_SpecificCondCode(ISD::SETNE))))
+      return Matched;
+
+    // Now try for bitcast
+    if (!sd_match(N, m_BitCast(m_Value(BitCast))))
+      return SDValue();
+
+    if (!sd_match(BitCast, m_c_SetCC(m_Value(Matched), m_Zero(),
+                                     m_SpecificCondCode(ISD::SETNE))))
+      return SDValue();
+    return Matched;
+  }
+
+  SmallVector<int> PowerIndices = buildMaskArrayByPower(Power, NumElements);
+  if (sd_match(N, m_And(m_Value(LHS),
+                        m_Shuffle(m_Value(LHS), m_VectorVT(m_Opc(ISD::POISON)),
+                                  m_SpecificMask(PowerIndices)))))
+    return matchAndOfShuffle(LHS.getNode(), Power + 1);
+
+  return SDValue();
+}
+static SDValue performExtractVecEltCombine(SDNode *N, SelectionDAG &DAG) {
+  using namespace llvm::SDPatternMatch;
+
+  assert(N->getOpcode() == ISD::EXTRACT_VECTOR_ELT);
+  SDLoc DL(N);
+
+  SDValue And;
+  if (!sd_match(N, m_ExtractElt(m_VectorVT(m_Value(And)), m_Zero())))
+    return SDValue();
+
+  if (SDValue Matched = matchAndOfShuffle(And.getNode(), 0))
+    return DAG.getZExtOrTrunc(
+        DAG.getNode(
+            ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
+            {DAG.getConstant(Intrinsic::wasm_alltrue, DL, MVT::i32), Matched}),
+        DL, N->getValueType(0));
+
+  return SDValue();
+}
 
 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG) {
   assert(N->getOpcode() == ISD::MUL);
@@ -3402,6 +3487,8 @@ WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N,
     return performTruncateCombine(N, DCI);
   case ISD::INTRINSIC_WO_CHAIN:
     return performLowerPartialReduction(N, DCI.DAG);
+  case ISD::EXTRACT_VECTOR_ELT:
+    return performExtractVecEltCombine(N, DCI.DAG);
   case ISD::MUL:
     return performMulCombine(N, DCI.DAG);
   }

llvm/test/CodeGen/WebAssembly/simd-reduceand.ll

@@ -6,18 +6,8 @@ define i1 @reduce_and_to_all_true_16i8(<16 x i8> %0) {
 ; CHECK-LABEL: reduce_and_to_all_true_16i8:
 ; CHECK:         .functype reduce_and_to_all_true_16i8 (v128) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    v128.const $push0=, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
-; CHECK-NEXT:    i8x16.ne $push10=, $0, $pop0
-; CHECK-NEXT:    local.tee $push9=, $0=, $pop10
-; CHECK-NEXT:    i8x16.shuffle $push1=, $0, $0, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push8=, $pop9, $pop1
-; CHECK-NEXT:    local.tee $push7=, $0=, $pop8
-; CHECK-NEXT:    i8x16.shuffle $push2=, $0, $0, 4, 5, 6, 7, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push3=, $pop7, $pop2
-; CHECK-NEXT:    i32x4.extract_lane $push4=, $pop3, 0
-; CHECK-NEXT:    i32.const $push5=, 0
-; CHECK-NEXT:    i32.ne $push6=, $pop4, $pop5
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i8x16.all_true $push0=, $0
+; CHECK-NEXT:    return $pop0
   %2 = icmp ne <16 x i8> %0, zeroinitializer
   %3 = sext <16 x i1> %2 to <16 x i8>
   %4 = bitcast <16 x i8> %3 to <4 x i32>
@@ -31,18 +21,8 @@ define i1 @reduce_and_to_all_true_4i32(<4 x i32> %0) {
 ; CHECK-LABEL: reduce_and_to_all_true_4i32:
 ; CHECK:         .functype reduce_and_to_all_true_4i32 (v128) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    v128.const $push0=, 0, 0, 0, 0
-; CHECK-NEXT:    i32x4.ne $push10=, $0, $pop0
-; CHECK-NEXT:    local.tee $push9=, $0=, $pop10
-; CHECK-NEXT:    i8x16.shuffle $push1=, $0, $0, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push8=, $pop9, $pop1
-; CHECK-NEXT:    local.tee $push7=, $0=, $pop8
-; CHECK-NEXT:    i8x16.shuffle $push2=, $0, $0, 4, 5, 6, 7, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push3=, $pop7, $pop2
-; CHECK-NEXT:    i32x4.extract_lane $push4=, $pop3, 0
-; CHECK-NEXT:    i32.const $push5=, 0
-; CHECK-NEXT:    i32.ne $push6=, $pop4, $pop5
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32x4.all_true $push0=, $0
+; CHECK-NEXT:    return $pop0
   %2 = icmp ne <4 x i32> %0, zeroinitializer
   %3 = sext <4 x i1> %2 to <4 x i32>
   %4 = tail call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> %3)
@@ -56,18 +36,8 @@ define i1 @reduce_and_to_all_true_2i64(<2 x i64> %0) {
 ; CHECK-LABEL: reduce_and_to_all_true_2i64:
 ; CHECK:         .functype reduce_and_to_all_true_2i64 (v128) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    v128.const $push0=, 0, 0, 0, 0
-; CHECK-NEXT:    i32x4.ne $push10=, $0, $pop0
-; CHECK-NEXT:    local.tee $push9=, $0=, $pop10
-; CHECK-NEXT:    i8x16.shuffle $push1=, $0, $0, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push8=, $pop9, $pop1
-; CHECK-NEXT:    local.tee $push7=, $0=, $pop8
-; CHECK-NEXT:    i8x16.shuffle $push2=, $0, $0, 4, 5, 6, 7, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    v128.and $push3=, $pop7, $pop2
-; CHECK-NEXT:    i32x4.extract_lane $push4=, $pop3, 0
-; CHECK-NEXT:    i32.const $push5=, 0
-; CHECK-NEXT:    i32.ne $push6=, $pop4, $pop5
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32x4.all_true $push0=, $0
+; CHECK-NEXT:    return $pop0
   %2 = bitcast <2 x i64> %0 to <4 x i32>
   %3 = icmp ne <4 x i32> %2, zeroinitializer
   %4 = sext <4 x i1> %3 to <4 x i32>