[llvm-branch-commits] [llvm] [LoongArch] Legalize BUILD_VECTOR into a broadcast when all non-undef elements are identical (PR #169755)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Mon Nov 16 01:16:32 PST 2026


https://github.com/zhaoqi5 updated https://github.com/llvm/llvm-project/pull/169755

>From 7d3d35e74a14511d999885ed847afa7c98281ab6 Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Thu, 27 Nov 2025 11:04:05 +0800
Subject: [PATCH 1/5] [LoongArch] Legalize BUILD_VECTOR into a broadcast when
 all non-undef elements are identical

When a BUILD_VECTOR consists of the same element (ignoring undefs),
it is better emitting a broadcast instead of multiple insertions.

Some floating-point cases suffer performance regressions, those
specific cases are excluded in this commit. Including when:

- only one element is non-undef,
- only two elements are non-undef, and one of them must at index 0,
- for v8f32 vector type, specially exclude the cases when the only
two non-undefs are at index (1,2)/(1,3)/(2,3).
---
 .../LoongArch/LoongArchISelLowering.cpp       | 36 ++++++++++++++++---
 1 file changed, 31 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index ab3d0c1a4488b6..9a89b77661db45 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -3944,6 +3944,36 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
   }
 
   if (!IsConstant) {
+    SmallVector<SDValue> Sequence;
+    BitVector UndefElements;
+    bool IsRepeated = Node->getRepeatedSequence(Sequence, &UndefElements);
+    unsigned SeqLen = Sequence.size();
+    unsigned NumUndefElts = UndefElements.count();
+
+    // When a BUILD_VECTOR consists of the same element (ignoring undefs),
+    // prefer emitting a broadcast instead of multiple insertions.
+    if (IsRepeated && SeqLen == 1) {
+      // Integer vectors benefit from splat unconditionally.
+      if (ResTy.isInteger())
+        return DAG.getSplatBuildVector(ResTy, DL, Sequence[0]);
+
+      // Only certain floating-point cases suffer performance regressions,
+      // exclude those specific cases.
+      bool IsSplatBetter = true;
+      if (NumUndefElts == NumElts - 1)
+        IsSplatBetter = false;
+      if (NumUndefElts == NumElts - 2 && !UndefElements[0])
+        IsSplatBetter = false;
+      if (ResTy == MVT::v8f32 && NumUndefElts == NumElts - 2 &&
+          ((!UndefElements[1] && !UndefElements[2]) ||
+           (!UndefElements[1] && !UndefElements[3]) ||
+           (!UndefElements[2] && !UndefElements[3])))
+        IsSplatBetter = false;
+
+      if (IsSplatBetter)
+        return DAG.getSplatBuildVector(ResTy, DL, Sequence[0]);
+    }
+
     // If the BUILD_VECTOR has a repeated pattern, use INSERT_VECTOR_ELT to fill
     // the sub-sequence of the vector and then broadcast the sub-sequence.
     //
@@ -3951,10 +3981,7 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
     // back to use INSERT_VECTOR_ELT to materialize the vector, because it
     // generates worse code in some cases. This could be further optimized
     // with more consideration.
-    SmallVector<SDValue> Sequence;
-    BitVector UndefElements;
-    if (Node->getRepeatedSequence(Sequence, &UndefElements) &&
-        UndefElements.count() == 0) {
+    if (IsRepeated && NumUndefElts == 0) {
       // Using LSX instructions to fill the sub-sequence of 256-bits vector,
       // because the high part can be simply treated as undef.
       SDValue Vector = DAG.getUNDEF(ResTy);
@@ -3966,7 +3993,6 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
 
       fillVector(Sequence, DAG, DL, Subtarget, FillVec, FillTy);
 
-      unsigned SeqLen = Sequence.size();
       unsigned SplatLen = NumElts / SeqLen;
       MVT SplatEltTy = MVT::getIntegerVT(VT.getScalarSizeInBits() * SeqLen);
       MVT SplatTy = MVT::getVectorVT(SplatEltTy, SplatLen);

>From c9a98cbdf22aafe4fcc2045ba2339669bf944797 Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Thu, 27 Nov 2025 11:13:06 +0800
Subject: [PATCH 2/5] update tests

---
 .../CodeGen/LoongArch/lasx/build-vector.ll    | 50 ++++---------------
 .../lasx/ir-instruction/insertelement.ll      | 10 ++--
 .../LoongArch/lasx/scalar-to-vector.ll        |  8 +--
 .../CodeGen/LoongArch/lsx/build-vector.ll     | 26 +++-------
 .../CodeGen/LoongArch/lsx/scalar-to-vector.ll |  8 +--
 5 files changed, 29 insertions(+), 73 deletions(-)

diff --git a/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll b/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
index f82189496bc50c..2209644de3db20 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/build-vector.ll
@@ -93,22 +93,7 @@ entry:
 define void @buildvector_v32i8_splat_with_undef(ptr %dst, i8 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v32i8_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 0
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 1
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 2
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 3
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 4
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 5
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 6
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 7
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 8
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 9
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 10
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 11
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 12
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 13
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 14
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 15
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -135,15 +120,7 @@ entry:
 define void @buildvector_v16i16_splat_with_undef(ptr %dst, i16 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v16i16_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 0
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 1
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 2
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 3
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 4
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 5
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 6
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 7
-; CHECK-NEXT:    xvpermi.q $xr0, $xr0, 2
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -162,9 +139,7 @@ entry:
 define void @buildvector_v8i32_splat_with_undef(ptr %dst, i32 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v8i32_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a1, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a1, 2
-; CHECK-NEXT:    xvpermi.q $xr0, $xr0, 2
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -189,8 +164,7 @@ define void @buildvector_v4i64_splat_with_undef(ptr %dst, i64 %a0) nounwind {
 ;
 ; LA64-LABEL: buildvector_v4i64_splat_with_undef:
 ; LA64:       # %bb.0: # %entry
-; LA64-NEXT:    xvinsgr2vr.d $xr0, $a1, 0
-; LA64-NEXT:    xvinsgr2vr.d $xr0, $a1, 3
+; LA64-NEXT:    xvreplgr2vr.d $xr0, $a1
 ; LA64-NEXT:    xvst $xr0, $a0, 0
 ; LA64-NEXT:    ret
 entry:
@@ -232,9 +206,8 @@ define void @buildvector_v8f32_splat_with_undef_2(ptr %dst, float %a0) nounwind
 ; CHECK-LABEL: buildvector_v8f32_splat_with_undef_2:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    # kill: def $f0 killed $f0 def $xr0
-; CHECK-NEXT:    xvinsve0.w $xr1, $xr0, 1
-; CHECK-NEXT:    xvinsve0.w $xr1, $xr0, 4
-; CHECK-NEXT:    xvst $xr1, $a0, 0
+; CHECK-NEXT:    xvreplve0.w $xr0, $xr0
+; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
   %ins1 = insertelement <8 x float> undef, float %a0, i32 1
@@ -246,11 +219,9 @@ entry:
 define void @buildvector_v8f32_splat_with_undef_3(ptr %dst, float %a0) nounwind {
 ; CHECK-LABEL: buildvector_v8f32_splat_with_undef_3:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    # kill: def $f0 killed $f0 def $vr0
-; CHECK-NEXT:    vextrins.w $vr1, $vr0, 16
-; CHECK-NEXT:    vextrins.w $vr1, $vr0, 48
-; CHECK-NEXT:    xvpermi.q $xr1, $xr1, 2
-; CHECK-NEXT:    xvst $xr1, $a0, 0
+; CHECK-NEXT:    # kill: def $f0 killed $f0 def $xr0
+; CHECK-NEXT:    xvreplve0.w $xr0, $xr0
+; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
   %ins1 = insertelement <8 x float> undef, float %a0, i32 1
@@ -292,8 +263,7 @@ define void @buildvector_v4f64_splat_with_undef_2(ptr %dst, double %a0) nounwind
 ; CHECK-LABEL: buildvector_v4f64_splat_with_undef_2:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    # kill: def $f0_64 killed $f0_64 def $xr0
-; CHECK-NEXT:    vextrins.d $vr0, $vr0, 16
-; CHECK-NEXT:    xvpermi.q $xr0, $xr0, 2
+; CHECK-NEXT:    xvreplve0.d $xr0, $xr0
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insertelement.ll b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insertelement.ll
index a13bdf9aeaf9f8..e6442fbbc77dd8 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insertelement.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insertelement.ll
@@ -35,7 +35,7 @@ define void @insert_32xi8_upper(ptr %src, ptr %dst, i8 %in) nounwind {
 define void @insert_32xi8_undef(ptr %dst, i8 %in) nounwind {
 ; CHECK-LABEL: insert_32xi8_undef:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 1
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = insertelement <32 x i8> poison, i8 %in, i32 1
@@ -46,8 +46,7 @@ define void @insert_32xi8_undef(ptr %dst, i8 %in) nounwind {
 define void @insert_32xi8_undef_upper(ptr %dst, i8 %in) nounwind {
 ; CHECK-LABEL: insert_32xi8_undef_upper:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 6
-; CHECK-NEXT:    xvpermi.q $xr0, $xr0, 2
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = insertelement <32 x i8> poison, i8 %in, i32 22
@@ -88,7 +87,7 @@ define void @insert_16xi16_upper(ptr %src, ptr %dst, i16 %in) nounwind {
 define void @insert_16xi16_undef(ptr %dst, i16 %in) nounwind {
 ; CHECK-LABEL: insert_16xi16_undef:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 1
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = insertelement <16 x i16> poison, i16 %in, i32 1
@@ -99,8 +98,7 @@ define void @insert_16xi16_undef(ptr %dst, i16 %in) nounwind {
 define void @insert_16xi16_undef_upper(ptr %dst, i16 %in) nounwind {
 ; CHECK-LABEL: insert_16xi16_undef_upper:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 2
-; CHECK-NEXT:    xvpermi.q $xr0, $xr0, 2
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a1
 ; CHECK-NEXT:    xvst $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = insertelement <16 x i16> poison, i16 %in, i32 10
diff --git a/llvm/test/CodeGen/LoongArch/lasx/scalar-to-vector.ll b/llvm/test/CodeGen/LoongArch/lasx/scalar-to-vector.ll
index bba269279937ab..39533cff7c868f 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/scalar-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/scalar-to-vector.ll
@@ -7,7 +7,7 @@
 define <32 x i8> @scalar_to_32xi8(i8 %val) {
 ; CHECK-LABEL: scalar_to_32xi8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.b $xr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <32 x i8> poison, i8 %val, i32 0
   ret <32 x i8> %ret
@@ -16,7 +16,7 @@ define <32 x i8> @scalar_to_32xi8(i8 %val) {
 define <16 x i16> @scalar_to_16xi16(i16 %val) {
 ; CHECK-LABEL: scalar_to_16xi16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <16 x i16> poison, i16 %val, i32 0
   ret <16 x i16> %ret
@@ -25,7 +25,7 @@ define <16 x i16> @scalar_to_16xi16(i16 %val) {
 define <8 x i32> @scalar_to_8xi32(i32 %val) {
 ; CHECK-LABEL: scalar_to_8xi32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <8 x i32> poison, i32 %val, i32 0
   ret <8 x i32> %ret
@@ -40,7 +40,7 @@ define <4 x i64> @scalar_to_4xi64(i64 %val) {
 ;
 ; LA64-LABEL: scalar_to_4xi64:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    vinsgr2vr.d $vr0, $a0, 0
+; LA64-NEXT:    xvreplgr2vr.d $xr0, $a0
 ; LA64-NEXT:    ret
   %ret = insertelement <4 x i64> poison, i64 %val, i32 0
   ret <4 x i64> %ret
diff --git a/llvm/test/CodeGen/LoongArch/lsx/build-vector.ll b/llvm/test/CodeGen/LoongArch/lsx/build-vector.ll
index 1fe289cdaaccb3..eb1a843bb42d84 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/build-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/build-vector.ll
@@ -93,14 +93,7 @@ entry:
 define void @buildvector_v16i8_splat_with_undef(ptr %dst, i8 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v16i8_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 0
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 2
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 4
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 6
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 8
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 10
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 12
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a1, 14
+; CHECK-NEXT:    vreplgr2vr.b $vr0, $a1
 ; CHECK-NEXT:    vst $vr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -119,10 +112,7 @@ entry:
 define void @buildvector_v8i16_splat_with_undef(ptr %dst, i16 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v8i16_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 1
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 3
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 5
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a1, 7
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a1
 ; CHECK-NEXT:    vst $vr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -137,8 +127,7 @@ entry:
 define void @buildvector_v4i32_splat_with_undef(ptr %dst, i32 %a0) nounwind {
 ; CHECK-LABEL: buildvector_v4i32_splat_with_undef:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a1, 1
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a1, 2
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a1
 ; CHECK-NEXT:    vst $vr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
@@ -158,7 +147,7 @@ define void @buildvector_v2i64_splat_with_undef(ptr %dst, i64 %a0) nounwind {
 ;
 ; LA64-LABEL: buildvector_v2i64_splat_with_undef:
 ; LA64:       # %bb.0: # %entry
-; LA64-NEXT:    vinsgr2vr.d $vr0, $a1, 0
+; LA64-NEXT:    vreplgr2vr.d $vr0, $a1
 ; LA64-NEXT:    vst $vr0, $a0, 0
 ; LA64-NEXT:    ret
 entry:
@@ -197,9 +186,8 @@ define void @buildvector_v4f32_splat_with_undef_2(ptr %dst, float %a0) nounwind
 ; CHECK-LABEL: buildvector_v4f32_splat_with_undef_2:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    # kill: def $f0 killed $f0 def $vr0
-; CHECK-NEXT:    vextrins.w $vr1, $vr0, 16
-; CHECK-NEXT:    vextrins.w $vr1, $vr0, 32
-; CHECK-NEXT:    vst $vr1, $a0, 0
+; CHECK-NEXT:    vreplvei.w $vr0, $vr0, 0
+; CHECK-NEXT:    vst $vr0, $a0, 0
 ; CHECK-NEXT:    ret
 entry:
   %ins1 = insertelement <4 x float> undef, float %a0, i32 1
@@ -969,7 +957,7 @@ define void @buildvector_v2i64_partial(ptr %dst, i64 %a0) nounwind {
 ;
 ; LA64-LABEL: buildvector_v2i64_partial:
 ; LA64:       # %bb.0: # %entry
-; LA64-NEXT:    vinsgr2vr.d $vr0, $a1, 0
+; LA64-NEXT:    vreplgr2vr.d $vr0, $a1
 ; LA64-NEXT:    vst $vr0, $a0, 0
 ; LA64-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/LoongArch/lsx/scalar-to-vector.ll b/llvm/test/CodeGen/LoongArch/lsx/scalar-to-vector.ll
index d2a506dd98547f..be1a8206c2418d 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/scalar-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/scalar-to-vector.ll
@@ -7,7 +7,7 @@
 define <16 x i8> @scalar_to_16xi8(i8 %val) {
 ; CHECK-LABEL: scalar_to_16xi8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.b $vr0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.b $vr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <16 x i8> poison, i8 %val, i32 0
   ret <16 x i8> %ret
@@ -16,7 +16,7 @@ define <16 x i8> @scalar_to_16xi8(i8 %val) {
 define <8 x i16> @scalar_to_8xi16(i16 %val) {
 ; CHECK-LABEL: scalar_to_8xi16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.h $vr0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <8 x i16> poison, i16 %val, i32 0
   ret <8 x i16> %ret
@@ -25,7 +25,7 @@ define <8 x i16> @scalar_to_8xi16(i16 %val) {
 define <4 x i32> @scalar_to_4xi32(i32 %val) {
 ; CHECK-LABEL: scalar_to_4xi32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a0
 ; CHECK-NEXT:    ret
   %ret = insertelement <4 x i32> poison, i32 %val, i32 0
   ret <4 x i32> %ret
@@ -40,7 +40,7 @@ define <2 x i64> @scalar_to_2xi64(i64 %val) {
 ;
 ; LA64-LABEL: scalar_to_2xi64:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    vinsgr2vr.d $vr0, $a0, 0
+; LA64-NEXT:    vreplgr2vr.d $vr0, $a0
 ; LA64-NEXT:    ret
   %ret = insertelement <2 x i64> poison, i64 %val, i32 0
   ret <2 x i64> %ret

>From 841a91dd3b75cbc7c1fc1fab0155db836a36bb7b Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Wed, 29 Apr 2026 17:12:04 +0800
Subject: [PATCH 3/5] update tests

---
 llvm/test/CodeGen/LoongArch/lasx/rotl-rotr.ll | 8 ++------
 llvm/test/CodeGen/LoongArch/lsx/rotl-rotr.ll  | 6 ++----
 2 files changed, 4 insertions(+), 10 deletions(-)

diff --git a/llvm/test/CodeGen/LoongArch/lasx/rotl-rotr.ll b/llvm/test/CodeGen/LoongArch/lasx/rotl-rotr.ll
index cde567366ad74a..12d5bf868a9740 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/rotl-rotr.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/rotl-rotr.ll
@@ -168,9 +168,7 @@ define void @rotl_v4i64(ptr %dst, ptr %src, i64 %a0) nounwind {
 ; LA32-LABEL: rotl_v4i64:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    xvld $xr0, $a1, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 2
-; LA32-NEXT:    xvpermi.q $xr1, $xr1, 2
+; LA32-NEXT:    xvreplgr2vr.w $xr1, $a2
 ; LA32-NEXT:    xvneg.d $xr1, $xr1
 ; LA32-NEXT:    xvrotr.d $xr0, $xr0, $xr1
 ; LA32-NEXT:    xvst $xr0, $a0, 0
@@ -199,9 +197,7 @@ define void @rotr_v4i64(ptr %dst, ptr %src, i64 %a0) nounwind {
 ; LA32-LABEL: rotr_v4i64:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    xvld $xr0, $a1, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 2
-; LA32-NEXT:    xvpermi.q $xr1, $xr1, 2
+; LA32-NEXT:    xvreplgr2vr.w $xr1, $a2
 ; LA32-NEXT:    xvrotr.d $xr0, $xr0, $xr1
 ; LA32-NEXT:    xvst $xr0, $a0, 0
 ; LA32-NEXT:    ret
diff --git a/llvm/test/CodeGen/LoongArch/lsx/rotl-rotr.ll b/llvm/test/CodeGen/LoongArch/lsx/rotl-rotr.ll
index 11c31af27211d2..709fdb800b0f5c 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/rotl-rotr.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/rotl-rotr.ll
@@ -168,8 +168,7 @@ define void @rotl_v2i64(ptr %dst, ptr %src, i64 %a0) nounwind {
 ; LA32-LABEL: rotl_v2i64:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    vld $vr0, $a1, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 2
+; LA32-NEXT:    vreplgr2vr.w $vr1, $a2
 ; LA32-NEXT:    vneg.d $vr1, $vr1
 ; LA32-NEXT:    vrotr.d $vr0, $vr0, $vr1
 ; LA32-NEXT:    vst $vr0, $a0, 0
@@ -198,8 +197,7 @@ define void @rotr_v2i64(ptr %dst, ptr %src, i64 %a0) nounwind {
 ; LA32-LABEL: rotr_v2i64:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    vld $vr0, $a1, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 0
-; LA32-NEXT:    vinsgr2vr.w $vr1, $a2, 2
+; LA32-NEXT:    vreplgr2vr.w $vr1, $a2
 ; LA32-NEXT:    vrotr.d $vr0, $vr0, $vr1
 ; LA32-NEXT:    vst $vr0, $a0, 0
 ; LA32-NEXT:    ret

>From 83d825c8688821b45437e7e1e3114a33f5e11d6d Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Fri, 13 Nov 2026 16:19:10 +0800
Subject: [PATCH 4/5] update after rebase

---
 .../LoongArch/lasx/scalar-load-to-vector.ll   | 22 +++++++++----------
 .../LoongArch/lsx/ir-instruction/sitofp.ll    |  4 ++--
 .../LoongArch/lsx/ir-instruction/uitofp.ll    |  4 ++--
 .../LoongArch/lsx/scalar-load-to-vector.ll    | 18 +++++++--------
 4 files changed, 24 insertions(+), 24 deletions(-)

diff --git a/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll b/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
index 4a750298e5ccca..5febff7eed471f 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
@@ -7,7 +7,7 @@
 define <32 x i8> @load_to_32xi8(ptr %p) {
 ; CHECK-LABEL: load_to_32xi8:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.b $vr0, $a0, 0
+; CHECK-NEXT:    xvldrepl.b $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i8, ptr %p
   %ret = insertelement <32 x i8> poison, i8 %v, i32 0
@@ -17,7 +17,7 @@ define <32 x i8> @load_to_32xi8(ptr %p) {
 define <16 x i16> @load_to_16xi16(ptr %p) {
 ; CHECK-LABEL: load_to_16xi16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vldrepl.h $vr0, $a0, 0
+; CHECK-NEXT:    xvldrepl.h $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i16, ptr %p
   %ret = insertelement <16 x i16> poison, i16 %v, i32 0
@@ -27,7 +27,7 @@ define <16 x i16> @load_to_16xi16(ptr %p) {
 define <8 x i32> @load_to_8xi32(ptr %p) {
 ; CHECK-LABEL: load_to_8xi32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fld.s $fa0, $a0, 0
+; CHECK-NEXT:    xvldrepl.w $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
   %ret = insertelement <8 x i32> poison, i32 %v, i32 0
@@ -44,7 +44,7 @@ define <4 x i64> @load_to_4xi64(ptr %p) {
 ;
 ; LA64-LABEL: load_to_4xi64:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    fld.d $fa0, $a0, 0
+; LA64-NEXT:    xvldrepl.d $xr0, $a0, 0
 ; LA64-NEXT:    ret
   %v = load i64, ptr %p
   %ret = insertelement <4 x i64> poison, i64 %v, i32 0
@@ -56,13 +56,13 @@ define <16 x i16> @load_to_16xi16_offset(ptr %p) {
 ; LA32-LABEL: load_to_16xi16_offset:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    addi.w $a0, $a0, -3
-; LA32-NEXT:    vldrepl.h $vr0, $a0, 0
+; LA32-NEXT:    xvldrepl.h $xr0, $a0, 0
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: load_to_16xi16_offset:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    addi.d $a0, $a0, -3
-; LA64-NEXT:    vldrepl.h $vr0, $a0, 0
+; LA64-NEXT:    xvldrepl.h $xr0, $a0, 0
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 -3
   %v = load i16, ptr %gep
@@ -80,7 +80,8 @@ define <4 x i64> @load_to_4xi64_offset(ptr %p) {
 ;
 ; LA64-LABEL: load_to_4xi64_offset:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    fld.d $fa0, $a0, 6
+; LA64-NEXT:    addi.d $a0, $a0, 6
+; LA64-NEXT:    xvldrepl.d $xr0, $a0, 0
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 6
   %v = load i64, ptr %gep
@@ -93,7 +94,7 @@ define <8 x i32> @load_to_8xi32_volatile(ptr %p) {
 ; CHECK-LABEL: load_to_8xi32_volatile:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a0
 ; CHECK-NEXT:    ret
   %v = load volatile i32, ptr %p
   %ret = insertelement <8 x i32> poison, i32 %v, i32 0
@@ -104,7 +105,7 @@ define <8 x i32> @load_to_8xi32_multi_use(ptr %p, ptr %dst) {
 ; CHECK-LABEL: load_to_8xi32_multi_use:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.w $xr0, $a0
 ; CHECK-NEXT:    st.w $a0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
@@ -116,8 +117,7 @@ define <8 x i32> @load_to_8xi32_multi_use(ptr %p, ptr %dst) {
 define <8 x i32> @load_to_8xi32_lane1(ptr %p) {
 ; CHECK-LABEL: load_to_8xi32_lane1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 1
+; CHECK-NEXT:    xvldrepl.w $xr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
   %ret = insertelement <8 x i32> poison, i32 %v, i32 1
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
index 33762d085815e9..68a5cf7626af7c 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
@@ -486,7 +486,7 @@ define <5 x double> @sitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a2, 1
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a3, 2
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a4, 3
-; LA32-NEXT:    fld.s $fa1, $a1, 16
+; LA32-NEXT:    vldrepl.w $vr1, $a1, 16
 ; LA32-NEXT:    vsllwil.d.w $vr2, $vr0, 0
 ; LA32-NEXT:    vffint.d.l $vr2, $vr2
 ; LA32-NEXT:    vexth.d.w $vr0, $vr0
@@ -507,7 +507,7 @@ define <5 x double> @sitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a2, 1
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a3, 2
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a4, 3
-; LA64-NEXT:    fld.s $fa1, $a1, 32
+; LA64-NEXT:    vldrepl.w $vr1, $a1, 32
 ; LA64-NEXT:    vsllwil.d.w $vr2, $vr0, 0
 ; LA64-NEXT:    vffint.d.l $vr2, $vr2
 ; LA64-NEXT:    vexth.d.w $vr0, $vr0
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
index 9ffe0a0bfd5638..6333093bef4eba 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
@@ -94,7 +94,7 @@ define <5 x double> @uitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA32-NEXT:    ld.w $a2, $a1, 4
 ; LA32-NEXT:    fld.s $fa0, $a1, 0
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a2, 1
-; LA32-NEXT:    fld.s $fa1, $a1, 16
+; LA32-NEXT:    vldrepl.w $vr1, $a1, 16
 ; LA32-NEXT:    ld.w $a2, $a1, 12
 ; LA32-NEXT:    fld.s $fa2, $a1, 8
 ; LA32-NEXT:    vrepli.b $vr3, 0
@@ -115,7 +115,7 @@ define <5 x double> @uitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA64-NEXT:    ld.w $a2, $a1, 8
 ; LA64-NEXT:    fld.s $fa0, $a1, 0
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a2, 1
-; LA64-NEXT:    fld.s $fa1, $a1, 32
+; LA64-NEXT:    vldrepl.w $vr1, $a1, 32
 ; LA64-NEXT:    ld.w $a2, $a1, 24
 ; LA64-NEXT:    fld.s $fa2, $a1, 16
 ; LA64-NEXT:    vrepli.b $vr3, 0
diff --git a/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll b/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
index d717a3c37647d2..7041f282a7a90a 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
@@ -27,7 +27,7 @@ define <8 x i16> @load_to_8xi16(ptr %p) {
 define <4 x i32> @load_to_4xi32(ptr %p) {
 ; CHECK-LABEL: load_to_4xi32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fld.s $fa0, $a0, 0
+; CHECK-NEXT:    vldrepl.w $vr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
   %ret = insertelement <4 x i32> poison, i32 %v, i32 0
@@ -44,7 +44,7 @@ define <2 x i64> @load_to_2xi64(ptr %p) {
 ;
 ; LA64-LABEL: load_to_2xi64:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    fld.d $fa0, $a0, 0
+; LA64-NEXT:    vldrepl.d $vr0, $a0, 0
 ; LA64-NEXT:    ret
   %v = load i64, ptr %p
   %ret = insertelement <2 x i64> poison, i64 %v, i32 0
@@ -80,7 +80,8 @@ define <2 x i64> @load_to_2xi64_offset(ptr %p) {
 ;
 ; LA64-LABEL: load_to_2xi64_offset:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    fld.d $fa0, $a0, 6
+; LA64-NEXT:    addi.d $a0, $a0, 6
+; LA64-NEXT:    vldrepl.d $vr0, $a0, 0
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 6
   %v = load i64, ptr %gep
@@ -155,7 +156,7 @@ define <4 x i32> @load_to_4xi32_volatile(ptr %p) {
 ; CHECK-LABEL: load_to_4xi32_volatile:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a0
 ; CHECK-NEXT:    ret
   %v = load volatile i32, ptr %p
   %ret = insertelement <4 x i32> poison, i32 %v, i32 0
@@ -166,7 +167,7 @@ define <4 x i32> @load_to_4xi32_multi_use(ptr %p, ptr %dst) {
 ; CHECK-LABEL: load_to_4xi32_multi_use:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.w $vr0, $a0
 ; CHECK-NEXT:    st.w $a0, $a1, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
@@ -178,8 +179,7 @@ define <4 x i32> @load_to_4xi32_multi_use(ptr %p, ptr %dst) {
 define <4 x i32> @load_to_4xi32_lane1(ptr %p) {
 ; CHECK-LABEL: load_to_4xi32_lane1:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    ld.w $a0, $a0, 0
-; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 1
+; CHECK-NEXT:    vldrepl.w $vr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p
   %ret = insertelement <4 x i32> poison, i32 %v, i32 1
@@ -201,7 +201,7 @@ define <4 x i32> @load_to_4xi32_not_undef(ptr %p, <4 x i32> %vec) {
 define <4 x i32> @load_to_4xi32_range(ptr %p) {
 ; CHECK-LABEL: load_to_4xi32_range:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    fld.s $fa0, $a0, 0
+; CHECK-NEXT:    vldrepl.w $vr0, $a0, 0
 ; CHECK-NEXT:    ret
   %v = load i32, ptr %p, !range !0
   %ret = insertelement <4 x i32> poison, i32 %v, i32 0
@@ -218,7 +218,7 @@ define <2 x i64> @load_to_2xi64_range(ptr %p) {
 ;
 ; LA64-LABEL: load_to_2xi64_range:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    fld.d $fa0, $a0, 0
+; LA64-NEXT:    vldrepl.d $vr0, $a0, 0
 ; LA64-NEXT:    ret
   %v = load i64, ptr %p, !range !1
   %ret = insertelement <2 x i64> poison, i64 %v, i32 0

>From a6d25ffb5e0a7f8649ad71291fe87bcf972be761 Mon Sep 17 00:00:00 2001
From: Qi Zhao <zhaoqi01 at loongson.cn>
Date: Mon, 16 Nov 2026 16:44:42 +0800
Subject: [PATCH 5/5] address heiher's comment

---
 .../LoongArch/LoongArchISelLowering.cpp       | 41 ++++++++++++-------
 .../LoongArch/lasx/scalar-load-to-vector.ll   |  7 ++--
 .../LoongArch/lsx/ir-instruction/sitofp.ll    |  4 +-
 .../LoongArch/lsx/ir-instruction/uitofp.ll    |  4 +-
 .../LoongArch/lsx/scalar-load-to-vector.ll    |  3 +-
 5 files changed, 34 insertions(+), 25 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 9a89b77661db45..7bfb6723fd377c 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -3953,22 +3953,33 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
     // When a BUILD_VECTOR consists of the same element (ignoring undefs),
     // prefer emitting a broadcast instead of multiple insertions.
     if (IsRepeated && SeqLen == 1) {
-      // Integer vectors benefit from splat unconditionally.
-      if (ResTy.isInteger())
-        return DAG.getSplatBuildVector(ResTy, DL, Sequence[0]);
-
-      // Only certain floating-point cases suffer performance regressions,
-      // exclude those specific cases.
       bool IsSplatBetter = true;
-      if (NumUndefElts == NumElts - 1)
-        IsSplatBetter = false;
-      if (NumUndefElts == NumElts - 2 && !UndefElements[0])
-        IsSplatBetter = false;
-      if (ResTy == MVT::v8f32 && NumUndefElts == NumElts - 2 &&
-          ((!UndefElements[1] && !UndefElements[2]) ||
-           (!UndefElements[1] && !UndefElements[3]) ||
-           (!UndefElements[2] && !UndefElements[3])))
-        IsSplatBetter = false;
+
+      // Integer vectors benefit from splat except in the case where the only
+      // non-undef element is at index 0 and is a load with a base+offset addr.
+      if (ResTy.isInteger()) {
+        SDValue Op0 = Node->getOperand(0);
+        if (NumUndefElts == NumElts - 1 && Op0.getOpcode() == ISD::LOAD &&
+            DAG.isBaseWithConstantOffset(Op0.getOperand(1)))
+          IsSplatBetter = false;
+      }
+
+      // Several floating-point cases suffer performance regressions or have
+      // no improvement when
+      //   only one element is non-undef,
+      //   one of the only two non-undefs at index 0,
+      //   the two non-undefs of v8f32 buildvector at index (1,2)/(1,3)/(2,3).
+      if (ResTy.isFloatingPoint()) {
+        if (NumUndefElts == NumElts - 1)
+          IsSplatBetter = false;
+        if (NumUndefElts == NumElts - 2 && !UndefElements[0])
+          IsSplatBetter = false;
+        if (ResTy == MVT::v8f32 && NumUndefElts == NumElts - 2 &&
+            ((!UndefElements[1] && !UndefElements[2]) ||
+             (!UndefElements[1] && !UndefElements[3]) ||
+             (!UndefElements[2] && !UndefElements[3])))
+          IsSplatBetter = false;
+      }
 
       if (IsSplatBetter)
         return DAG.getSplatBuildVector(ResTy, DL, Sequence[0]);
diff --git a/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll b/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
index 5febff7eed471f..aae8c58d488cc1 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/scalar-load-to-vector.ll
@@ -56,13 +56,13 @@ define <16 x i16> @load_to_16xi16_offset(ptr %p) {
 ; LA32-LABEL: load_to_16xi16_offset:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    addi.w $a0, $a0, -3
-; LA32-NEXT:    xvldrepl.h $xr0, $a0, 0
+; LA32-NEXT:    vldrepl.h $vr0, $a0, 0
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: load_to_16xi16_offset:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    addi.d $a0, $a0, -3
-; LA64-NEXT:    xvldrepl.h $xr0, $a0, 0
+; LA64-NEXT:    vldrepl.h $vr0, $a0, 0
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 -3
   %v = load i16, ptr %gep
@@ -80,8 +80,7 @@ define <4 x i64> @load_to_4xi64_offset(ptr %p) {
 ;
 ; LA64-LABEL: load_to_4xi64_offset:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    addi.d $a0, $a0, 6
-; LA64-NEXT:    xvldrepl.d $xr0, $a0, 0
+; LA64-NEXT:    fld.d $fa0, $a0, 6
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 6
   %v = load i64, ptr %gep
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
index 68a5cf7626af7c..33762d085815e9 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/sitofp.ll
@@ -486,7 +486,7 @@ define <5 x double> @sitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a2, 1
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a3, 2
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a4, 3
-; LA32-NEXT:    vldrepl.w $vr1, $a1, 16
+; LA32-NEXT:    fld.s $fa1, $a1, 16
 ; LA32-NEXT:    vsllwil.d.w $vr2, $vr0, 0
 ; LA32-NEXT:    vffint.d.l $vr2, $vr2
 ; LA32-NEXT:    vexth.d.w $vr0, $vr0
@@ -507,7 +507,7 @@ define <5 x double> @sitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a2, 1
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a3, 2
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a4, 3
-; LA64-NEXT:    vldrepl.w $vr1, $a1, 32
+; LA64-NEXT:    fld.s $fa1, $a1, 32
 ; LA64-NEXT:    vsllwil.d.w $vr2, $vr0, 0
 ; LA64-NEXT:    vffint.d.l $vr2, $vr2
 ; LA64-NEXT:    vexth.d.w $vr0, $vr0
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
index 6333093bef4eba..9ffe0a0bfd5638 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/uitofp.ll
@@ -94,7 +94,7 @@ define <5 x double> @uitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA32-NEXT:    ld.w $a2, $a1, 4
 ; LA32-NEXT:    fld.s $fa0, $a1, 0
 ; LA32-NEXT:    vinsgr2vr.w $vr0, $a2, 1
-; LA32-NEXT:    vldrepl.w $vr1, $a1, 16
+; LA32-NEXT:    fld.s $fa1, $a1, 16
 ; LA32-NEXT:    ld.w $a2, $a1, 12
 ; LA32-NEXT:    fld.s $fa2, $a1, 8
 ; LA32-NEXT:    vrepli.b $vr3, 0
@@ -115,7 +115,7 @@ define <5 x double> @uitofp_v5i32_v5f64(<5 x i32> %a) {
 ; LA64-NEXT:    ld.w $a2, $a1, 8
 ; LA64-NEXT:    fld.s $fa0, $a1, 0
 ; LA64-NEXT:    vinsgr2vr.w $vr0, $a2, 1
-; LA64-NEXT:    vldrepl.w $vr1, $a1, 32
+; LA64-NEXT:    fld.s $fa1, $a1, 32
 ; LA64-NEXT:    ld.w $a2, $a1, 24
 ; LA64-NEXT:    fld.s $fa2, $a1, 16
 ; LA64-NEXT:    vrepli.b $vr3, 0
diff --git a/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll b/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
index 7041f282a7a90a..b9296670ac328f 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/scalar-load-to-vector.ll
@@ -80,8 +80,7 @@ define <2 x i64> @load_to_2xi64_offset(ptr %p) {
 ;
 ; LA64-LABEL: load_to_2xi64_offset:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    addi.d $a0, $a0, 6
-; LA64-NEXT:    vldrepl.d $vr0, $a0, 0
+; LA64-NEXT:    fld.d $fa0, $a0, 6
 ; LA64-NEXT:    ret
   %gep = getelementptr i8, ptr %p, i64 6
   %v = load i64, ptr %gep



More information about the llvm-branch-commits mailing list