[llvm-branch-commits] [llvm] release/20.x: [CG][RISCV]Fix shuffling of odd number of input vectors (PR #125910)

[via llvm-branch-commits]

https://github.com/llvmbot created https://github.com/llvm/llvm-project/pull/125910

Backport 23b6a05ec9c2220844748487612761d1e09166b7

Requested by: @alexey-bataev

>From 37229643ffe8e98d53b7c3e8afe6d65dc305c927 Mon Sep 17 00:00:00 2001
From: Alexey Bataev <a.bataev at outlook.com>
Date: Wed, 5 Feb 2025 07:13:33 -0500
Subject: [PATCH] [CG][RISCV]Fix shuffling of odd number of input vectors

If the input contains odd number of shuffled vectors, the 2 last
shuffles are shuffled with the same first vector. Need to correctly
process such situation: when the first vector is requested for the first
time - extract it from the source vector, when it is requested the
second time - reuse previous result. The second vector should be
extracted in both cases.

Fixes #125269

Reviewers: topperc, preames

Reviewed By: preames

Pull Request: https://github.com/llvm/llvm-project/pull/125693

(cherry picked from commit 23b6a05ec9c2220844748487612761d1e09166b7)
---
 llvm/lib/Target/RISCV/RISCVISelLowering.cpp   | 24 +++++++++-------
 .../rvv/fixed-vectors-shuffle-exact-vlen.ll   | 28 +++++++++++++++++++
 2 files changed, 42 insertions(+), 10 deletions(-)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 8d09e534b1858b..7f0646fb10dfcc 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -5214,17 +5214,21 @@ static SDValue lowerShuffleViaVRegSplitting(ShuffleVectorSDNode *SVN,
     SmallDenseMap<unsigned, SDValue, 4> Values;
     for (unsigned I : seq<unsigned>(Data.size())) {
       const auto &[Idx1, Idx2, _] = Data[I];
-      if (Values.contains(Idx1)) {
-        assert(Idx2 != UINT_MAX && Values.contains(Idx2) &&
-               "Expected both indices to be extracted already.");
-        break;
+      // If the shuffle contains permutation of odd number of elements,
+      // Idx1 might be used already in the first iteration.
+      //
+      // Idx1 = shuffle Idx1, Idx2
+      // Idx1 = shuffle Idx1, Idx3
+      SDValue &V = Values.try_emplace(Idx1).first->getSecond();
+      if (!V)
+        V = ExtractValue(Idx1 >= NumOfSrcRegs ? V2 : V1,
+                         (Idx1 % NumOfSrcRegs) * NumOpElts);
+      if (Idx2 != UINT_MAX) {
+        SDValue &V = Values.try_emplace(Idx2).first->getSecond();
+        if (!V)
+          V = ExtractValue(Idx2 >= NumOfSrcRegs ? V2 : V1,
+                           (Idx2 % NumOfSrcRegs) * NumOpElts);
       }
-      SDValue V = ExtractValue(Idx1 >= NumOfSrcRegs ? V2 : V1,
-                               (Idx1 % NumOfSrcRegs) * NumOpElts);
-      Values[Idx1] = V;
-      if (Idx2 != UINT_MAX)
-        Values[Idx2] = ExtractValue(Idx2 >= NumOfSrcRegs ? V2 : V1,
-                                    (Idx2 % NumOfSrcRegs) * NumOpElts);
     }
     SDValue V;
     for (const auto &[Idx1, Idx2, Mask] : Data) {
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-exact-vlen.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-exact-vlen.ll
index afd560fd74d16a..c0c17d4e0623e7 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-exact-vlen.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-shuffle-exact-vlen.ll
@@ -431,3 +431,31 @@ define void @shuffle_i256_ldst(ptr %p) vscale_range(2,2) {
   store <4 x i256> %res, ptr %p
   ret void
 }
+
+define void @shuffle_3_input_vectors() vscale_range(4,4) {
+; CHECK-LABEL: shuffle_3_input_vectors:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e64, m8, ta, ma
+; CHECK-NEXT:    vmv.v.i v8, 1
+; CHECK-NEXT:    vsetivli zero, 1, e8, mf8, ta, ma
+; CHECK-NEXT:    vmv.v.i v0, 6
+; CHECK-NEXT:    vsetvli a0, zero, e64, m8, ta, ma
+; CHECK-NEXT:    vmv.v.i v16, 0
+; CHECK-NEXT:    vsetivli zero, 4, e64, m1, ta, mu
+; CHECK-NEXT:    vslidedown.vi v20, v8, 1, v0.t
+; CHECK-NEXT:    vslideup.vi v20, v9, 3
+; CHECK-NEXT:    vslidedown.vi v21, v9, 1
+; CHECK-NEXT:    vmv1r.v v22, v8
+; CHECK-NEXT:    vsetvli a0, zero, e64, m8, ta, ma
+; CHECK-NEXT:    vmsgt.vi v8, v16, 0
+; CHECK-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vmv.x.s a0, v8
+; CHECK-NEXT:    sb a0, 0(zero)
+; CHECK-NEXT:    ret
+  %1 = shufflevector <32 x i64> zeroinitializer, <32 x i64> splat (i64 1), <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 poison, i32 poison, i32 33, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+  %2 = icmp slt <32 x i64> zeroinitializer, %1
+  %3 = bitcast <32 x i1> %2 to i32
+  %4 = trunc i32 %3 to i8
+  store i8 %4, ptr null, align 1
+  ret void
+}