[SLP]Do not trunc bv nodes, if the user is vectorized an requires wider type.

If at least a single user of the gathered trunc'ed instruction is
vectorized and requires wider type, than the trunc node, such
gathers/buildvectors should not be optimized for better bitwidth.

Author: alexey-bataev

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -15529,7 +15529,23 @@ void BoUpSLP::computeMinimumValueSizes() {
     // Check if the root is trunc and the next node is gather/buildvector, then
     // keep trunc in scalars, which is free in most cases.
     if (E.isGather() && IsTruncRoot && E.UserTreeIndices.size() == 1 &&
-        E.Idx > (IsStoreOrInsertElt ? 2 : 1)) {
+        E.Idx > (IsStoreOrInsertElt ? 2 : 1) &&
+        all_of(E.Scalars, [&](Value *V) {
+          return V->hasOneUse() || isa<Constant>(V) ||
+                 (!V->hasNUsesOrMore(UsesLimit) &&
+                  none_of(V->users(), [&](User *U) {
+                    const TreeEntry *TE = getTreeEntry(U);
+                    const TreeEntry *UserTE = E.UserTreeIndices.back().UserTE;
+                    if (TE == UserTE || !TE)
+                      return false;
+                    unsigned UserTESz = DL->getTypeSizeInBits(
+                        UserTE->Scalars.front()->getType());
+                    auto It = MinBWs.find(TE);
+                    if (It != MinBWs.end() && It->second.first > UserTESz)
+                      return true;
+                    return DL->getTypeSizeInBits(U->getType()) > UserTESz;
+                  }));
+        })) {
       ToDemote.push_back(E.Idx);
       const TreeEntry *UserTE = E.UserTreeIndices.back().UserTE;
       auto It = MinBWs.find(UserTE);

llvm/test/Transforms/SLPVectorizer/RISCV/trunc-bv-multi-uses.ll

@@ -8,8 +8,7 @@ define i32 @test(i64 %v1, i64 %v2) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x i64> poison, i64 [[V1]], i32 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x i64> [[TMP0]], i64 [[V2]], i32 1
 ; CHECK-NEXT:    [[TMP2:%.*]] = trunc <2 x i64> [[TMP1]] to <2 x i32>
-; CHECK-NEXT:    [[TMP3:%.*]] = sext <2 x i32> [[TMP2]] to <2 x i64>
-; CHECK-NEXT:    [[TMP4:%.*]] = lshr <2 x i64> [[TMP3]], <i64 32, i64 32>
+; CHECK-NEXT:    [[TMP4:%.*]] = lshr <2 x i64> [[TMP1]], <i64 32, i64 32>
 ; CHECK-NEXT:    [[TMP5:%.*]] = trunc <2 x i64> [[TMP4]] to <2 x i32>
 ; CHECK-NEXT:    [[TMP6:%.*]] = add <2 x i32> [[TMP2]], [[TMP5]]
 ; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <2 x i32> [[TMP6]], i32 0