[LoongArch] Fix broadcast load with extension. (#155960)

PR #135896 introduces [x]vldrepl instructions without handling
extension.
This patch will fix that.

(cherry picked from commit 2320529)

Author: tangaac

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

@@ -2430,11 +2430,14 @@ static SDValue lowerBUILD_VECTORAsBroadCastLoad(BuildVectorSDNode *BVOp,
   }
 
   // make sure that this load is valid and only has one user.
-  if (!IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode()))
+  if (!IsIdeneity || !IdentitySrc || !BVOp->isOnlyUserOf(IdentitySrc.getNode()))
     return SDValue();
 
-  if (IsIdeneity) {
-    auto *LN = cast<LoadSDNode>(IdentitySrc);
+  auto *LN = cast<LoadSDNode>(IdentitySrc);
+  auto ExtType = LN->getExtensionType();
+
+  if ((ExtType == ISD::EXTLOAD || ExtType == ISD::NON_EXTLOAD) &&
+      VT.getScalarSizeInBits() == LN->getMemoryVT().getScalarSizeInBits()) {
     SDVTList Tys =
         LN->isIndexed()
             ? DAG.getVTList(VT, LN->getBasePtr().getValueType(), MVT::Other)

llvm/test/CodeGen/LoongArch/lasx/broadcast-load.ll

@@ -18,6 +18,32 @@ define <4 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst) {
   ret <4 x i64> %tmp2
 }
 
+define <16 x i16> @should_not_be_optimized_sext_load(ptr %ptr) {
+; CHECK-LABEL: should_not_be_optimized_sext_load:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ld.b $a0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a0
+; CHECK-NEXT:    ret
+  %tmp = load i8, ptr %ptr
+  %tmp1 = sext i8 %tmp to i16
+  %tmp2 = insertelement <16 x i16> zeroinitializer, i16 %tmp1, i32 0
+  %tmp3 = shufflevector <16 x i16> %tmp2, <16 x i16> poison, <16 x i32> zeroinitializer
+  ret <16 x i16> %tmp3
+}
+
+define <16 x i16> @should_not_be_optimized_zext_load(ptr %ptr) {
+; CHECK-LABEL: should_not_be_optimized_zext_load:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ld.bu $a0, $a0, 0
+; CHECK-NEXT:    xvreplgr2vr.h $xr0, $a0
+; CHECK-NEXT:    ret
+  %tmp = load i8, ptr %ptr
+  %tmp1 = zext i8 %tmp to i16
+  %tmp2 = insertelement <16 x i16> zeroinitializer, i16 %tmp1, i32 0
+  %tmp3 = shufflevector <16 x i16> %tmp2, <16 x i16> poison, <16 x i32> zeroinitializer
+  ret <16 x i16> %tmp3
+}
+
 define <4 x i64> @xvldrepl_d_unaligned_offset(ptr %ptr) {
 ; CHECK-LABEL: xvldrepl_d_unaligned_offset:
 ; CHECK:       # %bb.0:

llvm/test/CodeGen/LoongArch/lsx/broadcast-load.ll

@@ -1,8 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 ; RUN: llc --mtriple=loongarch64 -mattr=+lsx < %s | FileCheck %s
 
-; TODO: Load a element and splat it to a vector could be lowerd to vldrepl
-
 ; A load has more than one user shouldn't be lowered to vldrepl
 define <2 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst){
 ; CHECK-LABEL: should_not_be_optimized:
@@ -18,6 +16,32 @@ define <2 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst){
   ret <2 x i64> %tmp2
 }
 
+define <8 x i16> @should_not_be_optimized_sext_load(ptr %ptr) {
+; CHECK-LABEL: should_not_be_optimized_sext_load:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ld.b $a0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a0
+; CHECK-NEXT:    ret
+  %tmp = load i8, ptr %ptr
+  %tmp1 = sext i8 %tmp to i16
+  %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 0
+  %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer
+  ret <8 x i16> %tmp3
+}
+
+define <8 x i16> @should_not_be_optimized_zext_load(ptr %ptr) {
+; CHECK-LABEL: should_not_be_optimized_zext_load:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    ld.bu $a0, $a0, 0
+; CHECK-NEXT:    vreplgr2vr.h $vr0, $a0
+; CHECK-NEXT:    ret
+  %tmp = load i8, ptr %ptr
+  %tmp1 = zext i8 %tmp to i16
+  %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 0
+  %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer
+  ret <8 x i16> %tmp3
+}
+
 define <2 x i64> @vldrepl_d_unaligned_offset(ptr %ptr) {
 ; CHECK-LABEL: vldrepl_d_unaligned_offset:
 ; CHECK:       # %bb.0: