[SLP]Fix final analysis for unsigned nodes.

Need to check that at least single bit is cleared for unsigned nodes
before reducing their size. Otherwise they might be treated as signed in
signed nodes.

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -14442,11 +14442,18 @@ bool BoUpSLP::collectValuesToDemote(
     }
     auto NumSignBits = ComputeNumSignBits(V, *DL, 0, AC, nullptr, DT);
     unsigned BitWidth1 = OrigBitWidth - NumSignBits;
-    if (!isKnownNonNegative(V, SimplifyQuery(*DL)))
+    bool IsSigned = !isKnownNonNegative(V, SimplifyQuery(*DL));
+    if (IsSigned)
       ++BitWidth1;
     if (auto *I = dyn_cast<Instruction>(V)) {
       APInt Mask = DB->getDemandedBits(I);
       unsigned BitWidth2 = Mask.getBitWidth() - Mask.countl_zero();
+      while (!IsSigned && BitWidth2 < OrigBitWidth) {
+        APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth2 - 1);
+        if (MaskedValueIsZero(V, Mask, SimplifyQuery(*DL)))
+          break;
+        BitWidth2 *= 2;
+      }
       BitWidth1 = std::min(BitWidth1, BitWidth2);
     }
     BitWidth = std::max(BitWidth, BitWidth1);

llvm/test/Transforms/SLPVectorizer/RISCV/unsigned-node-trunc-with-signed-users.ll

@@ -8,18 +8,20 @@ define void @test(ptr %p, i16 %load794) {
 ; CHECK-NEXT:    [[GEP799:%.*]] = getelementptr inbounds i8, ptr [[P]], i64 16
 ; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x i16>, ptr [[P]], align 2
 ; CHECK-NEXT:    [[TMP2:%.*]] = load <2 x i16>, ptr [[GEP799]], align 2
-; CHECK-NEXT:    [[TMP3:%.*]] = sub <2 x i16> [[TMP2]], [[TMP1]]
-; CHECK-NEXT:    [[TMP4:%.*]] = add <2 x i16> [[TMP3]], <i16 3329, i16 3329>
+; CHECK-NEXT:    [[TMP3:%.*]] = zext <2 x i16> [[TMP1]] to <2 x i32>
+; CHECK-NEXT:    [[TMP4:%.*]] = zext <2 x i16> [[TMP2]] to <2 x i32>
+; CHECK-NEXT:    [[TMP7:%.*]] = sub nsw <2 x i32> [[TMP4]], [[TMP3]]
+; CHECK-NEXT:    [[TMP8:%.*]] = add nsw <2 x i32> [[TMP7]], <i32 3329, i32 3329>
 ; CHECK-NEXT:    [[TMP5:%.*]] = insertelement <2 x i32> poison, i32 [[ZEXT795]], i32 0
 ; CHECK-NEXT:    [[TMP6:%.*]] = shufflevector <2 x i32> [[TMP5]], <2 x i32> poison, <2 x i32> zeroinitializer
-; CHECK-NEXT:    [[TMP7:%.*]] = trunc <2 x i32> [[TMP6]] to <2 x i16>
-; CHECK-NEXT:    [[TMP8:%.*]] = mul <2 x i16> [[TMP4]], [[TMP7]]
-; CHECK-NEXT:    [[TMP9:%.*]] = zext <2 x i16> [[TMP8]] to <2 x i64>
+; CHECK-NEXT:    [[TMP12:%.*]] = mul <2 x i32> [[TMP8]], [[TMP6]]
+; CHECK-NEXT:    [[TMP9:%.*]] = zext <2 x i32> [[TMP12]] to <2 x i64>
 ; CHECK-NEXT:    [[TMP10:%.*]] = mul nuw nsw <2 x i64> [[TMP9]], <i64 5039, i64 5039>
 ; CHECK-NEXT:    [[TMP11:%.*]] = lshr <2 x i64> [[TMP10]], <i64 24, i64 24>
-; CHECK-NEXT:    [[TMP12:%.*]] = trunc <2 x i64> [[TMP11]] to <2 x i16>
-; CHECK-NEXT:    [[TMP13:%.*]] = mul <2 x i16> [[TMP12]], <i16 -3329, i16 -3329>
-; CHECK-NEXT:    [[TMP14:%.*]] = add <2 x i16> [[TMP13]], [[TMP8]]
+; CHECK-NEXT:    [[TMP13:%.*]] = trunc <2 x i64> [[TMP11]] to <2 x i32>
+; CHECK-NEXT:    [[TMP20:%.*]] = mul <2 x i32> [[TMP13]], <i32 62207, i32 62207>
+; CHECK-NEXT:    [[TMP21:%.*]] = add <2 x i32> [[TMP20]], [[TMP12]]
+; CHECK-NEXT:    [[TMP14:%.*]] = trunc <2 x i32> [[TMP21]] to <2 x i16>
 ; CHECK-NEXT:    [[TMP15:%.*]] = add <2 x i16> [[TMP14]], <i16 -3329, i16 -3329>
 ; CHECK-NEXT:    [[TMP16:%.*]] = icmp slt <2 x i16> [[TMP15]], zeroinitializer
 ; CHECK-NEXT:    [[TMP17:%.*]] = select <2 x i1> [[TMP16]], <2 x i16> [[TMP14]], <2 x i16> zeroinitializer