[llvm] 0397226 - [SLP]Fix PR90892: do a correct sign analysis of the entries elements in gather shuffles.

[Alexey Bataev via llvm-commits]

Author: Alexey Bataev
Date: 2024-05-03T14:01:25-07:00
New Revision: 03972261a93853d3e84857b4bed816648d3549f0

URL: https://github.com/llvm/llvm-project/commit/03972261a93853d3e84857b4bed816648d3549f0
DIFF: https://github.com/llvm/llvm-project/commit/03972261a93853d3e84857b4bed816648d3549f0.diff

LOG: [SLP]Fix PR90892: do a correct sign analysis of the entries elements in gather shuffles.

Need to do extra analysis of the scalar elements of the tree entry to be
shuffled instead of the vectorized value to correctly deduce signedness
info.

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/test/Transforms/SLPVectorizer/RISCV/shuffled-gather-casted.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index f3970f8b549c8c..43cf0ce6c2b926 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -11640,13 +11640,14 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
 
   /// Cast value \p V to the vector type with the same number of elements, but
   /// the base type \p ScalarTy.
-  Value *castToScalarTyElem(Value *V) {
+  Value *castToScalarTyElem(Value *V,
+                            std::optional<bool> IsSigned = std::nullopt) {
     auto *VecTy = cast<VectorType>(V->getType());
     if (VecTy->getElementType() == ScalarTy)
       return V;
     return Builder.CreateIntCast(
         V, VectorType::get(ScalarTy, VecTy->getElementCount()),
-        !isKnownNonNegative(V, SimplifyQuery(*R.DL)));
+        IsSigned ? *IsSigned : !isKnownNonNegative(V, SimplifyQuery(*R.DL)));
   }
 
 public:
@@ -11795,12 +11796,30 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
   /// Adds 2 input vectors (in form of tree entries) and the mask for their
   /// shuffling.
   void add(const TreeEntry &E1, const TreeEntry &E2, ArrayRef<int> Mask) {
-    add(E1.VectorizedValue, E2.VectorizedValue, Mask);
+    Value *V1 = E1.VectorizedValue;
+    if (V1->getType()->isIntOrIntVectorTy())
+      V1 = castToScalarTyElem(V1, all_of(E1.Scalars, [&](Value *V) {
+                                return !isKnownNonNegative(
+                                    V, SimplifyQuery(*R.DL));
+                              }));
+    Value *V2 = E2.VectorizedValue;
+    if (V2->getType()->isIntOrIntVectorTy())
+      V2 = castToScalarTyElem(V2, all_of(E2.Scalars, [&](Value *V) {
+                                return !isKnownNonNegative(
+                                    V, SimplifyQuery(*R.DL));
+                              }));
+    add(V1, V2, Mask);
   }
   /// Adds single input vector (in form of tree entry) and the mask for its
   /// shuffling.
   void add(const TreeEntry &E1, ArrayRef<int> Mask) {
-    add(E1.VectorizedValue, Mask);
+    Value *V1 = E1.VectorizedValue;
+    if (V1->getType()->isIntOrIntVectorTy())
+      V1 = castToScalarTyElem(V1, all_of(E1.Scalars, [&](Value *V) {
+                                return !isKnownNonNegative(
+                                    V, SimplifyQuery(*R.DL));
+                              }));
+    add(V1, Mask);
   }
   /// Adds 2 input vectors and the mask for their shuffling.
   void add(Value *V1, Value *V2, ArrayRef<int> Mask) {

diff  --git a/llvm/test/Transforms/SLPVectorizer/RISCV/shuffled-gather-casted.ll b/llvm/test/Transforms/SLPVectorizer/RISCV/shuffled-gather-casted.ll
index ded4b70966218d..c83d9363c1809c 100644
--- a/llvm/test/Transforms/SLPVectorizer/RISCV/shuffled-gather-casted.ll
+++ b/llvm/test/Transforms/SLPVectorizer/RISCV/shuffled-gather-casted.ll
@@ -9,7 +9,7 @@ define i32 @test(ptr %p) {
 ; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <4 x i16> <i16 0, i16 poison, i16 0, i16 0>, i16 [[D_0]], i32 1
 ; CHECK-NEXT:    [[TMP1:%.*]] = or <4 x i16> [[TMP0]], zeroinitializer
 ; CHECK-NEXT:    [[TMP2:%.*]] = and <4 x i16> [[TMP1]], zeroinitializer
-; CHECK-NEXT:    [[TMP3:%.*]] = sext <4 x i16> [[TMP0]] to <4 x i32>
+; CHECK-NEXT:    [[TMP3:%.*]] = zext <4 x i16> [[TMP0]] to <4 x i32>
 ; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <4 x i32> [[TMP3]], <4 x i32> <i32 0, i32 poison, i32 0, i32 0>, <4 x i32> <i32 4, i32 1, i32 6, i32 7>
 ; CHECK-NEXT:    [[TMP5:%.*]] = icmp sgt <4 x i32> [[TMP4]], zeroinitializer
 ; CHECK-NEXT:    [[TMP6:%.*]] = select <4 x i1> [[TMP5]], <4 x i16> [[TMP2]], <4 x i16> <i16 0, i16 2, i16 0, i16 0>