[llvm] 913dcf1 - [SLP]Fix type promotion for smax reduction with unsigned reduced operands

[Alexey Bataev via llvm-commits]

Author: Alexey Bataev
Date: 2025-04-16T10:14:29-07:00
New Revision: 913dcf1aa36f3ea2d67a0d2b05b9d1375987e553

URL: https://github.com/llvm/llvm-project/commit/913dcf1aa36f3ea2d67a0d2b05b9d1375987e553
DIFF: https://github.com/llvm/llvm-project/commit/913dcf1aa36f3ea2d67a0d2b05b9d1375987e553.diff

LOG: [SLP]Fix type promotion for smax reduction with unsigned reduced operands

Need to add an extra bit for sign info for unsigned reduced values to
generate correct code.

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
    llvm/test/Transforms/SLPVectorizer/RISCV/smax-reduction-unsigned-missing-sign.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 810d44343c4a9..83252bdb51ea2 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -20521,7 +20521,8 @@ void BoUpSLP::computeMinimumValueSizes() {
   }
   bool IsSignedCmp = false;
   if (UserIgnoreList && all_of(*UserIgnoreList, [](Value *V) {
-        return match(V, m_SMin(m_Value(), m_Value()));
+        return match(V, m_SMin(m_Value(), m_Value())) ||
+               match(V, m_SMax(m_Value(), m_Value()));
       }))
     IsSignedCmp = true;
   while (NodeIdx < VectorizableTree.size()) {

diff  --git a/llvm/test/Transforms/SLPVectorizer/RISCV/smax-reduction-unsigned-missing-sign.ll b/llvm/test/Transforms/SLPVectorizer/RISCV/smax-reduction-unsigned-missing-sign.ll
index e6408572acf8f..a8efc2622aa80 100644
--- a/llvm/test/Transforms/SLPVectorizer/RISCV/smax-reduction-unsigned-missing-sign.ll
+++ b/llvm/test/Transforms/SLPVectorizer/RISCV/smax-reduction-unsigned-missing-sign.ll
@@ -7,8 +7,9 @@ define i32 @test(i8 %0) {
 ; CHECK-NEXT:  [[ENTRY:.*:]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <4 x i8> <i8 poison, i8 0, i8 0, i8 0>, i8 [[TMP0]], i32 0
 ; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne <4 x i8> [[TMP1]], zeroinitializer
-; CHECK-NEXT:    [[TMP3:%.*]] = call i1 @llvm.vector.reduce.smax.v4i1(<4 x i1> [[TMP2]])
-; CHECK-NEXT:    [[TMP4:%.*]] = zext i1 [[TMP3]] to i32
+; CHECK-NEXT:    [[TMP3:%.*]] = zext <4 x i1> [[TMP2]] to <4 x i8>
+; CHECK-NEXT:    [[TMP5:%.*]] = call i8 @llvm.vector.reduce.smax.v4i8(<4 x i8> [[TMP3]])
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i8 [[TMP5]] to i32
 ; CHECK-NEXT:    ret i32 [[TMP4]]
 ;
 entry: